US 5,594,885 AGrant

Method for Operating a Cache Memory System Using a Recycled Register for Identifying a Reuse Status of a Corresponding Cache Entry

Issue Date:1997-01-14

•16 Claims

•21 Drawing Sheets

Abstract

A computer data storage device made up of both solid state storage and rotating magnetic disk which maintains a fast response time approaching that of a solid state device for many workloads and improving on the response time of a normal magnetic disk for practically all workloads. The high performance is accomplished by a special hardware configuration coupled with unique procedures and algorithms pertaining to the methodology of placing and maintaining data in the most appropriate media based on actual and projected activity. The system management features a completely searchless method for determining the location of data within and between the two devices. Sufficient solid state memory capacity is incorporated to permit retention of useful, active data, as well as to permit prefetching of data into the solid state component when the probabilities favor such action. Movement of updated data from the solid state storage to the magnetic disk and of prefetched data from the magnetic disk to the solid state component is done on a timely, but unobtrusive, basis as background tasks of the described device. The direct, private channel between the solid state storage and the magnetic disk prevents the conversations between these two media from conflicting with the transmission of data between the host computer and the described device. A set of microprocessors manage and oversee the data transmission and storage. Data integrity is maintained through a power interruption via a battery assisted, automatic and intelligent shutdown procedure.

Metadata

Assignee

Zitel Corporation

Inventor

Marvin Lautzenheiser

Application Information

Application Number:US 2552515

Filing Date:1994-06-07

Priority Date:1991-03-05

Art Unit:232

Classifications

IPC:

G06F 1212

Field of Search:

395403;440;445;463;461;470;487;460;486;471

Patent Drawings (21 sheets)

Description

Introduction

1. Field of Invention

This invention relates to a high performance computer data storage device including a combination of solid state storage and a rotating magnetic disk device.

2. Description of Prior Art

A number of computer data storage systems exist which make some use of solid state memory devices as a caching controller for placement between a computer host device and rotating magnetic disk devices. A typical caching system uses a single solid state memory unit as a holding area for data stored on a string of magnetic disks, thereby allowing certain information to be stored in a high speed cache memory, thereby increasing speed of performance as compared to the use solely of relatively lower speed disk memories, i.e. the percentage of times a desired piece of data is contained in the high speed cache memory, thereby allowing faster access as compared with when that data is only stored in a disk drive. A block diagram of such a system is shown in FIG. 1. Host computer 101 communicates with the entire string 102 of disks 102-1 through 102-N via cache unit 103 via Host interface 104, such as Small computer Systems Interface (SCSI). All data going to or from disk strip 102 passes through the cache-to-disk data path consisting of host interface 104, cache unit 103, and disk interface 105. Cache unit 103 manages the caching of data and services requests from host computer 101. Major components of cache unit 103 include microprocessor 103-1, cache management hardware 103-2, cache management firmware 103-3, address lookup table 103-4, and solid state cache memory 103-5.

The prior art cache system of FIG. 1 is intended to hold frequently accessed data in a solid state memory area so as to give more rapid access to that data than would be achieved if the same data were accessed from the disk media. Typically, such cache systems are quite effective when attached to certain host computers and under certain workloads. However, there exist some drawbacks and, under certain conditions, such cache systems exhibit a performance level less than that achieved by similar, but uncached, devices. Some of the factors contributing to the less than desirable performance of prior art cached disk devices are now described.

The single cache memory 103-5 is used in conjunction with all disks in disk string 102. Data from any of the disks may reside in cache memory 103-5 at any given time. The currently accessed data is given precedence for caching regardless of the disk drive on which it resides. When fulfilling a host command, the determination of whether or not the data is in cache memory 103-5, and the location of that data in cache memory 103-5, is usually via hashing schemes and table search operations. Hashing schemes and table searches can introduce time delays of their own which can defeat the purpose of the cache unit itself.

Performance is very sensitive to cache-hit rates. Due to caching overhead and queuing times, a low hit rate in a typical string oriented cache system can result in overall performance that is poorer than that of configured uncached string of disks.

The size of cache memory 103-5 relative to the capacity of disk drives 102 is generally low. An apparently obvious technique to remedy a low hit rate is to increase the cache memory 103-5 size. However, it has been found that there is an upper limit to the size of cache memory 103-5 above which adding more capacity has limited benefits. With limited cache memory 103-5 capacity, a multitude of requests over a variety of data segments exhausts the capability of the cache system to retain the desirable data in cache memory 103-5. Often, data that would be reused in the near future is decached prematurely to make room in cache memory 103-5 for handling new requests from the host computer 101. The result is a reduced cache hit rate. A reduced hit rate increases the number of disk accesses; increased disk accesses increases the contention on the data path. A self-defeating cycle is instituted.

"Background" cache-ahead operations are limited since the data transferred during such activities travels over the same data path as, and often conflicts with, the data transferred to service direct requests from the host computer 101. The data path between cache unit 103 and disk string 102 can easily be overloaded. All data to and from any of the disks in disk string 102, whether for satisfying requests from host computer 101 or for cache management purposes, travels across the cache-to-disk path. This creates a bottleneck if a large amount of prefetching of data from disk string 102 to cache memory 103-5 occurs. Each attempt to prefetch data from disk string 102 into cache memory 103-5 potentially creates contention for the path with data being communicated between any of the disk drives of disk string 102 and host computer 101. As a result, prefetching of data into cache memory 103-5 must be judiciously limited; increasing the size of the cache memory 103-5 beyond a certain limit does not produce corresponding improvements in the performance of the cache system. This initiates a string of related phenomena. Cache-ahead management is often limited to fetching an extra succeeding track of data from disk wherever a read command from the host cannot be fulfilled from the cached data. This technique helps to minimize the tendency of cache-ahead to increase the queuing of requests waiting for the path between cache memory 103-5 and disk string 102. However, one of the concepts on which caching is based is that data accesses tend to be concentrated within a given locality within a reasonably short time frame. For example, data segments are often accessed in sequential fashion. Limiting the cache-ahead operations to being a function of read misses can have the negative effect of lowering the cache hit rate since such limitation may prevent or degrade the exploitation of the locality of data accesses.

A variety of algorithms and configurations have been devised in attempts to optimize the performance of string caches. A nearly universally accepted concept involves the retention and replacement of cached data segments based on least-recently used (LRU) measurements. The decaching of data to make room for new data is managed by a table which gives, for each cached block of data, its relative time since it was last accessed. Depending on the algorithm used, this process can also result in some form of table search with a potential measurable time delay.

Cache memory 103-5 is generally volatile; the data is lost if power to the unit is removed. This characteristic, coupled with the possibility of unexpected power outages, has generally imposed a write-through design for handling data transferred from host computer 103 to the cached string. In such a design, all writes from host computer 103 are written directly to disk; handled at disk speed, these operations are subject to all the inherent time delays of seek, latency, and lower transfer rates commonly associated with disk operations.

Cache unit 103 communicates with the string of disk drives 102 through disk interface 105.

Summary of the Invention

Computer operations and throughput are often limited by the time required to write data to, or read data from, a peripheral data storage device. A solid state storage device has high-speed response, but at a relatively high cost per megabyte of storage. A rotating magnetic disk, optical disk, or other mass media provides high storage capacity at a relatively low cost per megabyte, but with a low-speed response. The teachings of this invention provide a hybrid solid state and mass storage device which gives near solid state speed at a cost per megabyte approaching that of the mass storage device.

For the purposes of this discussion, embodiments will be described with regard to magnetic disk media. However, it is to be understood that the teachings of this invention are equally applicable to other types of mass storage devices, including optical disk devices, and the like.

This invention is based on a combination of hardware and firmware features.

The hardware features include: one or more rotating magnetic disk media, an ample solid state storage capacity; private channels between the disks and the solid state storage device; and high speed microprocessors to gather the intelligence, make data management decisions, and carry out the various data management asks.

The firmware features include the logic for gathering the historical data, making management decisions, and instructing the hardware to carry out the various data management operations. Important aspects of the firmware include making the decisions regarding the retention of data in the solid state memory based on usage history gathered during the device's work load experience.

The present invention includes a unique methodology for retaining, or recycling, of certain cached data which normally would be decached. While it would be normal to decache that data which is least recently used, this invention adds the further feature of utilizing a simple, but effective method of determining the probability of reuse of the least recently used data. This recycling methodology determines which data, although currently the least recently used, should still be retained in cache based on its higher potential reuse; and which least recently used data has a lessor probability of being reused, and thus, should be decached to make space in cache for other data which is or may be of current need.

The hybrid storage media of this invention performs at near solid state speeds for many types of computer workloads while practically never performing at less than normal magnetic disk speeds for any workload.

A rotating magnetic disk media is used to give the device a large capacity; the solid state storage is used to give the device a high-speed response capability. By associating the solid state media directly with a single magnetic disk device, a private data communication line is established which avoids contention between normal data transfers between the host and the device and transfers between the solid state memory and the disk. This private data channel permits virtually unlimited conversation between the two storage media. Utilization of ample solid state memory permits efficient maintenance of data for multiple, simultaneously active data streams. Management of the storage is via one or more microprocessors which utilize historical and projected data accesses to perform intelligent placement of data. No table searches are employed in the time-critical path. Host accesses to data stored in the solid state memory are at solid state speeds; host accesses to data stored on the magnetic disk are at disk device speeds. Under most conditions, all data sent from the host to the device is handled at solid state speeds.

Brief Descriptions of the Drawings

FIG. 1 is a block diagram of a typical prior art cached disk computer data storage system;

FIG. 2 is a block diagram depicting one embodiment of a cached disk computer data storage device constructed in accordance with the teachings of this invention;

FIG. 3 is a block diagram depicting one embodiment of a hardware controller for implementing the described invention;

FIG. 4 is a flow chart depicting the operation of one embodiment of this invention;

FIG. 5 is a flow chart depicting a more detailed description of the operation of the host command step of FIG. 4;

FIG. 6 is a flow chart depicting the operation of one embodiment of the analyze host I/O command operation of FIG. 5;

FIG. 7 is a flow chart depicting in more detail the operation of the setup track address list operation of FIG. 6;

FIG. 8 is a flow chart depicting in more detail the address translation of FIG. 7;

FIG. 9 is a flow chart depicting the cache read hit operation depicted in FIG. 5;

FIG. 10 is a flow chart depicting in more detail the cache read miss/operation depicted in FIG. 5;

FIG. 11 is a flow chart depicting the cache write hit operation of FIG. 5;

FIG. 12 is a flow chart depicting the cache write miss operation of FIG. 5;

FIG. 13 is a flow chart depicting the seek cache miss operation of FIG. 5;

FIG. 14 is a flow chart depicting the decache LRU operation of FIGS. 6, 13 and 15;

FIG. 15 is a flow chart depicting the cache ahead operation of FIG. 4;

FIG. 16 is a flow chart depicting the operation of the cache ahead determination operation of FIG. 15;

FIG. 17 is a flow chart depicting the operation of the initiate background sweep operation of FIG. 4;

FIG. 18 is a flow chart depicting the step of background sweep initiation at host I/O completion depicted in FIG. 4;

FIG. 19 is a flow chart depicting the generate background event operations depicted in FIGS. 17, 18, and 20;

FIG. 20 is a flow chart depicting the operation of the continued background sweep step of FIG. 4;

FIG. 21 is a flow chart depicting the power down control operations; and

FIG. 22 is a flow chart depicting the final background sweep operation depicted in FIG. 21.

Description of the Tables

Tables F-1 through F-4 describe the organization of Tables T-1 through T-4, respectively;

Table T-0 depicts a sample of I/O commands extracted from computer system operations during normal usage. These I/O commands, and the intervening commands, were the basis for the sample predicted LRU and ADT tables as shown in Tables T-1 through T-3.

Table T-1 depicts an example of values in the address translation (ADT) table prior to the handling of the first I/O operation from the host CPU;

Table T-2 depicts an example of values in the Least-Recently-Used (LRU) table prior to the handling of the first I/O operation from the host CPU; and

Table 3 is formed of Tables T-3a through T-3e, which depict the ADT table after various numbers of I/O operations.

Detailed Description of the Invention

Glossary of Terms

ADDRESS TRANSLATION:

The conversion of a sector address into a track address and sector offset within the track. CACHE-AHEAD FACTOR; PROXIMITY FACTOR:

At each track hit or rehit, cached data sufficient to satisfy a number of I/O's may remain in front of, and/or behind, the current location of the data involved in the current I/O. When either of these two remaining areas contain valid data for less than a set number of I/O's, the cache-ahead is activated. That minimum number of potential I/O's is the cache-ahead factor, or the proximity factor.

ADDRESS TRANSLATION TABLE; ADT TABLE:

The table which maintains the relationship between disk track identifiers and solid state memory addresses; also may hold frequency of access and/or other information as required.

CACHE:

The solid state memory area which holds user data within the cache system of this invention.

CPU SECTOR:

See Logical Sector.

DISK; MAGNETIC DISK; ROTATING MAGNETIC DISK:

A rotating magnetic media disk drive.

DISK SECTOR ADDRESS:

The address of a physical sector on the magnetic disk device.

DISK SERVER:

The logical section of the caching device which handles the writes to, and reads from, the rotating magnetic disk.

DISK TRACK ADDRESS; TRACK ADDRESS:

The address of the first sector of data in a given track on disk. These addresses correspond to physical locations on the rotating magnetic disk. Each sector address as specified in an I/O operation can be converted into a track address and a sector offset within that track.

DMA:

Direct Memory Access; that is, memory-to-memory transfer without the involvement of the processor.

DRAM:

Dynamic random access memory. The chip or chips that are used for solid state memory devices.

EDAC:

Error Detection And Correction

EEPROM:

Electrically Erasable Programmable Read-Only Memory

EPROM:

Erasable Programmable Read-Only Memory

HOST:

The computer to which the caching device is attached.

HOST SERVER:

The portion of the caching device which interfaces with the host computer.

I/O SIZE:

The size of a host I/O request as a number of sectors.

LOGICAL BLOCK:

See TRACK.

LOGICAL SEGMENT:

One or more contiguous sectors within a logical track.

LOGICAL TRACK:

See TRACK.

LRU:

Least Recently Used, describes the data currently occupying a cache data storage track and which has not been accessed for the longest period of time of all currently cached data. This is a well known concept for determining which cached data to release from a cache track in order to be able to reuse the cache space currently occupied by that data for caching some currently required, uncached data.

LRU TABLE; LEAST-RECENTLY-USED TABLE:

The table containing the information which allows the caching device's controller to determine which solid state memory data areas may be reused with the least impact on the cache efficiency.

MASS STORAGE DEVICE:

A rotating magnetic disk, optical disk, or other mass media which provides high storage capacity at a relatively low cost per megabyte, but with a low-speed response.

MODIFIED DATA:

That data stored in the cache which has been written from a host to this described device and which has not yet been written by this described device to the mass storage device.

MODIFIED SEGMENT:

One or more contiguous sectors within a logical block which contain data written from the host to the cache and which data has not been subsequently written to the mass storage device.

MRU:

Most-Recently-Used, as pertains to that data storage track which has been accessed in the nearest time past.

NORMAL MODE:

The condition of the device in which it can use its normal priorities in order to reach its optimal performance level.

NULL, NULL VALUE:

A value in a table field which indicates the field should be considered to be empty; depending on usage, will be zero, or will be the highest value the bit structure of the field can accommodate.

PHYSICAL TRACK; DISK TRACK:

A complete data track on a disk; one complete band on one platter of the disk device.

PROXIMITY FACTOR:

See Cache-Ahead Factor.

READ-MISS-MAXSIZE:

The size of a host read transaction as a number of sectors which, when exceeded, causes the transaction to be handled in pseudo disk mode.

RECYCLE:

The term used to describe the retention of data in a track in cache beyond that tracks arrival at the LRU position; such retention may be based on a number of factors, including whether or not the track was used at some time since the data in the track was most recently read from disk into cache, or since the cached data track was last retained in cache by the recycling mechanism.

RECYCLE FLAG:

SEE RECYCLE REGISTER.

RECYCLING MECHANISM:

The term used to describe an entire set of procedures whose function it is to retain in cache data beyond the time that data would have been retained had the retention been based solely on the standard LRU concept. The recycling mechanism maintains and uses in decisions a recycle flag or recycle register.

RECYCLE REGISTER:

The term used to describe a register or data field, one of which is associated with each cache track, and whose value is adjusted based on the activity of the data cached in that track. The value in the recycle register is used to help make the decisions as to which cache tracks to be reused when a cache track is required for caching a currently uncached track. In its simplest form, this can be a single bit, and can be considered simply as a RECYCLE FLAG which is set to one when the data in a cache track qualifies for recycling, and is set to zero when the data in the cache track no longer qualifies for recycling.

SCSI:

Small Computer System Interface; the name applied to the protocol for interfacing devices, such as a disk device to a host computer.

SCSI CONTROL CHANNEL:

A physical connection between devices which uses the SCSI protocol, and is made up of logical controllers connected by a cable.

SECTOR:

The logical sub-unit of a disk track; the smallest addressable unit of data on a disk.

SOLID STATE MEMORY, SOLID STATE DEVICE; SSD:

Storage media made up of solid state devices such as DRAMs.

SSD TRACK ADDRESS:

The address in the solid state memory at which the first byte of the first sector of a given disk track resides.

TRACK; LOGICAL TRACK; LOGICAL BLOCK:

A logical data track on disk, or its equivalent in SSD; may or may not be identical to a physical track on disk (one complete magnetic band on one platter of the disk). It is noted that an I/O operation may involve more than one logical block.

TRACK SIZE:

The number of sectors considered to be in a disk track; this may or may not be equal to the actual number of sectors in a physical disk track.

URGENT MODE:

The condition of the device in which it must shift priorities in order to maintain at least magnetic disk level performance.

WRITE-MISS-MAXSIZE:

The size of a host write transaction as a number of sectors which, when exceeded, causes the transaction to be handled in pseudo disk mode.

System Overview

In accordance with the teachings of this invention, a computer peripheral data storage device is provided comprising a combination solid state memory and rotating magnetic disk; such device having the large capacity of magnetic disk media with near solid state speed at a cost per megabyte approaching that of magnetic disk media. For the purposes of this discussion, embodiments will be described with regard to magnetic disk media. However, it is to be understood that the teachings of this invention are equally applicable to other types of mass storage devices, including optical disk devices, and the like.

The caching device described herein derives its large storage capacity from the rotating magnetic disk media. Its high speed performance stems from the combination of a private channel between the two storage media, multiple microprocessors utilizing a set of unique data management algorithms, a unique prefetch procedure, combined in a methodology which incorporates simultaneity of memory management and data storage operations and an ample solid state memory. This hybrid storage media gives overall performance near that of solid state memory for most types of computer workloads while practically never performing at less than normal magnetic disk speeds for any workload.

To the host computer, the device of this invention appears to be a single, directly addressable entity. By the combination, within the device, of a solid state memory and one or more magnetic disk devices, private data communication lines are established within the device which avoids contention between normal data transfers between the host and the device, and transfers between the solid state memory and the disk media. This private data channel permits unrestricted data transfers between the two storage media with practically no contention with the communication between the host computer and the described device. Utilization of ample solid state memory permits efficient retention of data for multiple, simultaneously active data streams. Management of the storage is via microprocessors which anticipate data accesses based on historical activity. Data is moved into the solid state memory from the disk media based on management algorithms which insure that no table searches need be employed in the time-critical path. Host computer accesses to data stored in the solid state memory are at near solid state speeds; accesses to data stored on the magnetic disk are at near disk device speeds. All data sent from the host to the device is transferred at solid state speeds limited only by the channel capability.

Hardware Description

A device constructed in accordance with the teachings of this invention is depicted in FIG. 2. Memory device 200 is a self-contained module which includes interfaces with certain external devices. Its primary contact is with host computer 201 via host interface 204. Host interface 204 comprises, for example, a dedicated SCSI control processor which handles communications between host computer 201 and memory manager 205. An operator interface is provided via the console 207, which allows the user to interrogate as well as exercise overall control of the memory device 200. Another method of interfacing with the caching device 200 is by means of dial-in line 202 operating through the console.

Memory manager 205 handles all functions necessary to manage the storage of data in, and retrieval of data from disk drive 210 (or high capacity memory devices) and solid state memory 208, the two storage media. The memory manager 205 consists of one or more microprocessors associated firmware 205-2, and management tables, such as Address Translation (ADT) Table 205-3 and Least Recently Used (LRU) Table 205-4.

Solid state memory 208 is utilized for that data which memory manager 205, based on its experience, deems most useful to host computer 201, or most likely to become useful in the near future.

Magnetic disk 201 is the ultimate storage for all data, and provides the needed large storage capacity. Disk interface 209 serves as a separate dedicated control processor (such as an SCSI processor) for handling communications between memory manager 205 and disk drive 210.

Information about functional errors and operational statistics are maintained by diagnostic module-error logger 206. Access to module 206 is obtained through console 207. Console 207 serves as the operator's access to the memory device 200 for such actions as reading or resetting the error logger, or inquiring of the caching device's status or operating statistics.

The memory device 200 includes power backup system 203 which includes a rechargeable battery. Backup system 203 is prepared to maintain power to memory device 200 should normal power be interrupted. If such a power interruption occurs, the memory manager 205 takes whatever action is necessary to place all updated data stored in solid state memory 208 onto magnetic disk 210 before shutting down memory device 200.

FIG. 3 depicts a hardware controller block diagram of one embodiment of this invention. As shown in FIG. 3, hardware controller 300 provides three I/O ports, 301, 302, and 303. I/O ports 301 and 302 are differential SCSI ports used to connect hardware controller 300 to one or more host computers 201 (FIG. 2). I/O port 303 is a single-ended SCSI port used to connect controller 300 to disk drive 210 (which in this embodiment is a 5.25" magnetic hard disk drive). Disk drive 210 provides long-term non-volatile storage for data that flows into controller 300 from host computers 201. "Differential" and "single-ended" refer to specific electrical characteristics of SCSI ports; the most significant distinction between the two lies in the area of acceptable I/O cable length. The SCSI aspects of I/O ports 301, 302, and 303 are otherwise identical.

Cache memory 308 (corresponding to memory 208) is a large, high-speed memory used to store, on a dynamic basis, the currently active and potentially active data. The storage capacity of cache memory 308 can be selected at any convenient size and, in the embodiment depicted in FIG. 3, comprises 64 Megabytes of storage. Cache memory 308 is organized as 16 Megawords; each word consists of four data bytes (32 bits) and seven bits of error-correcting code. Typically, the storage capacity of cache memory 308 is selected to be within the range of approximately one-half of one percent (0.5) to 100 percent of the storage capacity of the one or more magnetic disks 210 (FIG. 2) with which it operates. A small portion of cache memory 308 is used to store the tables required to manage the caching operations; alternatively, a different memory (not shown, but accessible by microcontroller 305) is used for this purpose.

EDAC circuitry 306 performs error detecting and correcting functions for cache memory 308. In this embodiment, EDAC circuitry 306 generates a seven-bit error-correcting code for each 32-bit data word written to cache memory 308; this information is written to cache memory 308 along with the data word from which it was generated. The error-correcting code is examined by EDAC circuitry 306 when data is retrieved from cache memory 308 to verify that the data has not been corrupted since last written to cache memory 308. The modified Hamming code chosen for this embodiment allows EDAC circuitry 306 to correct all single-bit errors that occur and detect all double-bit and many multiple-bit errors that occur.

Error logger 307 is used to provide a record of errors that are detected by EDAC circuitry 306. The information recorded by error logger 307 is retrieved by microcontroller 305 for analysis and/or display. This information is sufficiently detailed to permit identification by microcontroller 305 of the specific bit in error (for single-bit errors) or the specific word in error (for double-bit errors). In the event that EDAC circuitry 306 detects a single-bit error, the bit in error is corrected as the data is transferred to whichever interface requested the data (processor/cache interface logic 316, host/cache interface logic 311 or 312, and disk/cache interface logic 313). A signal is also sent to microcontroller 305 to permit handling of this error condition (which involves analyzing the error based on the contents of error logger 307, attempting to scrub (correct) the error, and analyzing the results of the scrub to determine if the error was soft or hard).

In the event that EDAC circuitry 306 detects a double-bit error, a signal is sent to microcontroller 305. Microcontroller 305 will recognize that some data has been corrupted. If the corruption has occurred in the ADT or LRU tables, an attempt is made to reconstruct the now-defective table from the other, then relocate both tables to a different portion of cache memory 308.

If the corruption has occurred in an area of cache memory 308 that holds user data, microcontroller 305 attempts to salvage as much data as possible (transferring appropriate portions of cache memory 308 to disk drive 210, for example) before refusing to accept new data transfer commands. Any response to a request for status from the host computer 201 will contain information that the host computer 201 may use to recognize that memory device 200 is no longer operating properly.

Microcontroller 305 includes programmable control processor 314 (for example, an 80C196 microcontroller available from Intel Corporation of Santa Clara, Calif.), 64 kilobytes of EPROM memory 315, and hardware to allow programmable control processor 314 to control the following: I/O ports 301, 302, and 303, cache memory 308, EDAC 306, error logger 307, host/cache interface logic 311 and 312, disk/cache interface logic 313, processor/cache interface logic 316, and serial port 309.

Programmable control processor 314 performs the functions dictated by software programs that have been converted into a form that it can execute directly. These software programs are stored in EPROM memory 315.

In one embodiment, the host/cache interface logic sections 311 and 312 are essentially identical. Each host/cache interface logic section contains the DMA, byte/word, word/byte, and address register hardware that is required for the corresponding I/O port (301 for 311, 302 for 312) to gain access to cache memory 308. Each host/cache interface logic section also contains hardware to permit control via microcontroller 305. In this embodiment I/O ports 301 and 302 have data path widths of eight bits (byte). Cache memory 308 has a data path width of 32 bits (word).

Disk/cache interface logic 313 is similar to host/cache interface logic sections 311 and 312. It contains the DMA, byte/word, word/byte, and address register hardware that is required for disk I/O port 303 to gain access to cache memory 308. Disk/cache interface logic 313 also contains hardware to permit control via microcontroller 305. In this embodiment, I/O port 303 has a data path width of eight bits (byte).

Processor/cache interface logic 316 is similar to host/cache interface logic sections 311 and 312 and disk/cache interface logic 313. It contains the DMA, half-word/word, word/half-word, and address register hardware that is required for programmable control processor 314 to gain access to cache memory 308. Processor/cache interface logic 316 also contains hardware to permit control via microcontroller 305. In this embodiment, programmable control processor 314 has a data path width of 16 bits (half-word).

Serial port 309 allows the connection of an external device (for example, a small computer) to provide a human interface to the system 200. Serial port 309 permits initiation of diagnostics, reporting of diagnostic results, setup of system 200 operating parameters, monitoring of system 200 performance, and reviewing errors recorded inside system 200. In other embodiments, serial port 309 allows the transfer of different and/or improved software programs from the external device to the control program storage (when memory 315 is implemented with EEPROM rather than EPROM, for example).

Formats of control tables

Format of Address Translation (ADT) Table

The Address Translation Table, along with the LRU table, maintains the information required to manage the caching operations. There are two sections in the ADT table, the indexed, or tabular portion, and the set of unindexed, or single-valued items.

The unindexed portion of the ADT table contains two types of data fields; the first are those items which are essential to the cache management, the second category contains those data items which maintain records of the unit's performance.

The first group of unindexed items, or those requisite to the cache management, includes the following single-valued items.

1) ADT-CNL.

The number of tracks on the cached disk spindle; also equals the number of lines in the ADT table. This is set at the time the caching device is configured and is not changed while the unit is in operation.

2) ADT-HEAD-POS.

The current position of the read/write head of the cache disk. This is updated every time the head is positioned.

3) ADT-SWEEP-DIR.

The direction in which the current sweep of the background writes is progressing. This is updated each time the sweep reverses its direction across the disk.

4) ADT-MOD-COUNT.

The total number of tracks in the cache which have been modified by writes from the host and are currently awaiting a write to disk by the Disk server. This is increased by one whenever an unmodified cache track is updated by the host, and it is decreased by one whenever a modified cache track is copied to the cache disk.

ADT-MOD-URGENT.

The number of cache slots which, when in the modified condition, causes the caching device to shift priorities to maintain optimal performance.

The second group of unindexed items are those which record the unit's performance, and are all used to compute the current operating characteristics of the unit. They include the following single-valued items.

1) ADT-READ-HITS.

The number of cache read-hits encountered since the last reset. This value is set to zero by a reset operation from the console. It is incremented by one for each read I/O which is entirely satisfied from data which is resident in the cache memory.

2) ADT-READ-MISSES.

The number of cache read-misses encountered since the last reset. This value is set to zero by a reset operation from the console. It is incremented by one for each read I/O which cannot be entirely satisfied from data which is resident in the cache memory.

3) ADT-WRITE-HITS.

The number of cache write-hits encountered since the last reset. This value is set to zero by a reset operation from the console. It is incremented by one for each write I/O for which the corresponding track or tracks are found to be in cache memory.

4) ADT-WRITE-MISSES.

The number of cache write-misses encountered since the last reset. This value is set to zero by a reset operation from the console. It is incremented by one for which at least one of the corresponding track is not found to be in cache memory.

There is one line in the tabular portion for each data track on the spindle. A line is referred to by its line number, or index. That line number directly corresponds to a track number on the disk. When the host wants to access or modify data on the disk, it does so by referencing a starting sector address and indicating the number of sectors to be accessed or modified. For caching purposes, the starting sector address is converted into a track identifier and offset within that track.

A disk sector address is converted into a track number and a sector offset by dividing it by the number of sectors per track. The remainder is the offset into the track. The quotient is the track identifier and is the index into the ADT table. Using this index, the condition of the specified disk track can be determined directly from data in the ADT table; no search is required to determine cache-hits or misses.

Each ADT line contains the following items:

1) ADT-SLOT.

The number of the cache slot which contains the data for the disk track corresponding to this ADT table line number. By design, the value in ADT-SLOT also points to the line in the LRU table related to the cached disk track. If the disk track is not in cache memory, the value in this field is meaningless and is set to its null value. It is by means of this field that cache-hits can be serviced completely without any table search. A null value in this field indicates the corresponding disk track is not stored in the cache. This field is updated each time a track is entered into or removed from the SSD area.

2) ADT-MODIFIED.

A flag indicating whether or not the corresponding cached track has been modified by a write operation from the host, and thus, needs to be copied from the cache to the disk.

Format of Least Recently Used (LRU) Table

The LRU table maintains the information relative to the times when cached tracks of data were last accessed. This information is necessary for the unit to always be aware of which cache slots are available for overwriting whenever uncached data tracks must be placed in cache. Its contents also provide redundancy for the data kept in the ADT table, thus contributing to system reliability.

There are two sections in the LRU table, the indexed, or tabular portion, and the set of unindexed, or single-valued items. The unindexed portion of the LRU table contains data required to manage the caching process. The tabular portion is composed of pointers for LRU chaining purposes, pointers into the ADT table, and the recycle control registers or flags.

It is by means of this LRU information and the ADT table information that the system determines which cached track to overwrite when a cache area is needed for an uncached disk track. The unindexed items are requisite to the cache management, and includes the following single-valued items.

1) LRU-CNL.

The number of track-equivalent slots in the cache area; this is equal to the number of lines in the LRU table.

2) LRU-LRU.

The LRU-LRU table element points to the cache area track-slot containing the cached data which has been left un-touched for the longest time. It is updated when new activity for the referenced slot makes it no longer the least-recently-used. The referenced slot is the top candidate for overwriting when new data must be written into the cache.

3) LRU-MRU.

The LRU-MRU table element points to the cache area track-slot containing the cached data which has been most-recently referenced by the host. LRU-MRU is updated every time a track is touched by either a read or a write from the host. At that time, the address of the accessed track is placed in LRU-MRU and the LRU chains are updated in the indexed portion of the LRU table.

There is one line in the tabular portion for each cache area slot in the cache data area. A line is referred to by its line number, or index. That line number directly corresponds to a slot in the cache data area.

Each LRU table line contains pointer fields plus other control fields.

1) LRU-TRACK.

The pointer to the ADT line which references the disk track currently resident in the corresponding cache slot. By design, this value is also the identifier of the disk track whose data currently resides in the corresponding cache slot, if any.

2) LRU-LAST.

This is part of the bidirectional chaining of the cache data slots. LRU-LAST is the pointer to the next-older (in usage) cache slot. If this slot is the oldest, LRU-LAST will contain a zero.

3) LRU-NEXT.

This is the other half of the bidirectional chaining of the cache data slots. LRU-NEXT is the pointer to the next newer (in usage) cache slot. If this slot is the newest, LRU-NEXT will contain a zero.

4) LRU-CACHED-LOW.

A field containing the track-relative number of the lowest sector of this cached track which contains valid cached data.

5) LRU-CACHED-HIGH.

A field containing the track-relative number of the highest sector of this cached track which contains valid cached data.

6) LRU-MOD-LOW.

A field containing the track-relative number of the lowest sector of this cached track which contains modified cached data.

7) LRU-MOD-HIGH.

A field containing the track-relative number of the highest sector of this cached track which contains modified cached data.

8) LRU-LOCKED.

A flag indicating whether or not the corresponding cached track is currently the target of some operation, such as being acquired from the disk, being modified by the host, or being written to the disk by the cache controller; such operation making the track unavailable for certain other operations.

9) LRU-RECYCLE-REGISTER.

This field is used to control the recycling mechanism. It is increased or reduced based on the usage of the data cached in the corresponding track and based on other system factors such as the amount of modified data in the entire cache at various relevant times. This register's adjusted value is used to determine whether to decache or recycle the data in this cache track when this track arrives at a decision point such as when it reaches the LRU position in the LRU table. In its simplest form, this register becomes a single-bit RECYCLE-FLAG. In this simplified case, a single bit marker is maintained to indicate whether or not the corresponding track should be recycled. This flag is set to 1 (on) whenever the data in the track is referenced by the host; the flag is set to 0 (off) when the corresponding track is recycled (moved to the MRU position). The flag is initially set to 0 when a track is brought into cache as a result of a cache-ahead decision. For a track brought in to satisfy a cache miss, it is set to 1. In the more complete form, the recycle register can assume values from zero to n and the value is controlled by a set of recycling rules. See recycling mechanism examples elsewhere in this document.

Examples of Tables

Initial ADT table

When a unit is first powered on, the ADT table is in an indeterminate state. In order to become operational, initial values must be entered into their appropriate table elements. Initial values for unindexed fields of the ADT table are as follows:

The ADT-CNL field must be set to the size of the cache disk as a number of tracks.

The ADT-HEAD-POS field is set to zero to indicate the head is currently at the edge of the disk. This may, or may not, be true, but it does not matter; it will become correct on the first access to the disk.

The ADT-SWEEP-DIR field is arbitrarily set to one (1) to indicate the head is moving in an upward (based on track addresses) direction. This will be corrected at the initiation of the first background sweep.

The ADT-MOD-COUNT field is set to zero to reflect the fact that no modified tracks are waiting in cache to be copied to disk.

The ADT-READ-HITS field is set to zero to reflect the fact that no cache hits have occurred during read operations.

The ADT-READ-MISSES field is set to zero to reflect the fact that no cache misses have occurred during read operations.

The ADT-WRITE-HITS field is set to zero to reflect the fact that no cache hits have occurred during write operations.

The ADT-WRITE-MISSES field is set to zero to reflect the fact that no cache misses have occurred during write operations.

All indexed fields of all lines of the ADT table are initially set to zero to indicate that no tracks are resident in cache.

Initial LRU table

When the described caching device is first powered on, the LRU table is in an indeterminate state. In order to become operational, initial values must be entered into their appropriate table elements. While there are many acceptable ways to initialize the chaining fields, a simple one has been selected, and is described here.

Initial values for unindexed fields of the LRU table are as follows:

The LRU-CNL field must be set to the size of the cache, as a number of track-equivalent slots.

The LRU-LRU field is set to one to represent the lowest numbered cache slot as being the oldest. This is an arbitrary choice in keeping with the chaining values selected, below.

The LRU-MRU field is set equal to LRU-CNL to represent the highest cache slot as being the most recently used. This is an arbitrary choice in keeping with the initial chaining values selected, below.

Initial values for indexed fields of the LRU table are as follows:

The LRU-TRACK field of every line of the LRU table is set to zero to indicate that no disk data tracks are currently held in cache.

The LRU-LAST field of every line of the LRU table is set to that line's index minus one. This action, along with the settings for the LRU-NEXT values, produce a chained list suitable for the cache start-up operation.

The LRU-NEXT field of every line, except the highest, of the LRU table is set to that line's index plus one. The LRU-NEXT field of the highest line is set to zero. These settings, along with the settings for the LRU-LAST values, produces a chained list suitable for the cache start-up operation.

The LRU-CACHED-LOW field of every line is set to its null value to indicate that no portion of the disk track is currently held in cache.

The LRU-CACHED-HIGH field of every line is set to its null value to indicate that no portion of the disk track is currently held in cache.

The LRU-MOD-LOW field of every line is set to its null value to indicate that no portion of the disk track currently held in cache is in a modified condition.

The LRU-MOD-HIGH field of every line is set to its null value to indicate that no portion of the disk track currently held in cache is in a modified condition.

The LRU-LOCKED field of every line of the LRU table is set to zero to indicate no cache slot is currently locked.

The LRU-RECYCLE-REGISTER field of every line of the LRU table is set to zero to indicate that no slot is currently a candidate for recycling. While any set of recycling algorithms could have been used, the following rules have been chosen for this example:

1) The recycle register will be a single bit;

2) Cache tracks whose data is originally cached due to a cache read miss will have their recycling register set to one at the time of their original caching. This will give that data a longer life in cache than data cached for some of the other reasons;

3) Cache tracks whose data is originally cached by prefetch (based on proximity) will have their recycling register set to zero. This will give that data the minimum time in cache unless the data is referenced by the host before it is decached when it reaches the LRU position.

4) Cache tracks whose data is originally cached by a cache write miss will have their recycling register set to zero. Again, data cached in this manner will be given a minimum time in cache unless the data is referenced by the host before it is decached when it reaches the LRU position.

5) Cache tracks whose data is subsequently referenced by a host operation will have their recycle register set to one.

Operational state ADT tables

The operational state ADT table examples illustrate the conditions after the very first (sample) I/O has occurred and after the system has reached a fully active condition. These fully active examples show the effects of several I/O's on the state of the ADT table contents. Also included in these examples is the effect of a background sweep which wrote modified tracks from cache to disk. A detailed description of these specific sample operations appears under the LRU table discussion, below.

Operational state LRU tables

The operational state LRU table examples illustrate the conditions after the very first I/O has occurred and after the system has reached a fully active condition. These fully active examples show the effects of several I/O's on the state of the LRU table contents.

Description of Sample I/O's

For purposes of illustration, a sample of I/O's were chosen to be discussed in detail. Those chosen for discussion are I/O numbers 1, 1000, 1001, and 1002 taken from a trace of actions at an operational site; they are detailed in Table T-0 as projected into the described system. The following discussions of the sample I/O's include the effects on both the ADT and the LRU tables.

Actions related to I/O operation number 1:

1. This I/O is a read involving disk track numbers 46 and 47; since nothing is in cache, it must be a cache-miss. A portion of track 46, and all of track 47 is brought into cache. The ADT table is modified to show the locations of the tracks in cache; the amount of each track now cached is recorded; the chains are relinked to show the tracks to be the MRU and the next-to-MRU tracks; and they are both marked for recycling. They are marked for recycling since the data in these cache tracks were cached due to a read miss. According to the recycling rules set out for this example, cache tracks containing data brought from the disk as a result of a cache read miss are to be given a recycle register value of one.

2. Based on the I/O size and the distance from the end of the data requested in the I/O operation number 1 to the end of track 47, a prefetch of track 48 is initiated. That activity is not reflected in the ADT and LRU tables since it is initiated as a background operation after the completion of the current I/O.

After 999 I/O's have occurred, the ADT and LRU tables have reached a certain status. I/O number 1000 is a read of 68 sectors starting at sector address 14,190. This occupies sectors 11 through 78 of disk track 56. Based on these conditions, the following actions relate to I/O operation number 1000:

1. This I/O is a read involving track 56 which is not in cache; the required portion of it must be brought into cache. While fetching this required data, the portion of track 56 from the end of the requested data to the end of the track is also fetched. This is done here since this is the most expeditious time to do so, and satisfies the prefetch rules. The LRU table is updated to reflect the caching of track 56 and slot into which this data is fetched is placed at the MRU position in the LRU chain.

2. To make room for caching track 56, the old LRU track was decached.

3. A read I/O operation does not affect the need for a background sweep. There are three tracks in cache that need to be copied to disk; this condition remains unchanged by the current I/O.

4. Several cache tracks have non-zero recycling registers, including the track at the LRU end of the chain and the track it points to. Before any prefetch is initiated, these tracks will be moved to the MRU and next-to-MRU positions and their recycle registers will have been set to zero. This is in accordance with the recycle rules chosen for the current example wherein the recycle register is a single bit which is either 1) a one to indicate the data in the cache track is to be retained beyond its arrival at the LRU position; or 2) a zero which indicates the data in the cache track should be decached when it arrives at the LRU position and a cache track is needed for data from some other disk track.

5. Since track 57 is already in cache, no prefetch is needed for it.

5. The size of the I/O (68 sectors) and the current I/O's proximity to the first sector in track 56 indicate that track 55 should be prefetched by a cache-ahead action. That prefetch will be initiated as a background operation. Based on the LRU chain and the recycling situation, track 55 will be cached into slot 10. For the moment it will occupy the MRU position in the chain.

Actions related to I/O operation number 1001:

1. Prefetching of track 55, which was initiated by I/O 1000 has now been completed.

2. I/O number 1001 is a write involving sectors 191-256 of track 61, and sectors 1-2 of track 62. The LRU and ADT table references to them are updated.

3. This I/O action modified two cached tracks, bringing the total number of tracks which need written to disk up to the trigger point for the background sweep.

4. The background sweep is initiated and starts writing the modified tracks from cache to disk.

5. Since the background sweep is using the disk spindle, no cache-ahead is initiated, even though the unit would consider the prefetch of track 60 into cache.

Actions related to I/O operation number 1002:

1. The background sweep completed writing all modified tracks from cache to disk; it then went into the dormant state.

2. I/O 1002 was a write involving track 214 which is already resident in cache. The track is marked in the ADT table as having been modified. In the LRU table, track 214 in slot number 13 is removed from the MRU-LRU chain, and the amount modified is recorded in the LRU-MOD-LOW and LRU-MOD-HIGH fields.

3. A prefetch of track 215 is initiated since the position of the current I/O in track 214 is near enough to the end of the track to warrant a cache-ahead operation. This activity does not appear in the ADT and LRU tables for I/O 1002 since it will occur in the background after the completion of the current I/O.

4. Since a prefetch of track 215 has been initiated, track 213 is not considered for prefetching.

Firmware

Firmware Overview

The memory controller of this invention goes into a semi-dormant state when there is no activity that it needs to handle. As depicted in the flow chart of FIG. 4, there are three types of occurrences that may cause the controller to become active:

1. The host computer sends a command;

2. The background sweep completes an event;

3. The background sweep times out.

Insofar as possible, the host computer commands are given priority over other memory device activities. Thus, when a command is received from the host, it is immediately turned over to the Host Command Handler (described elsewhere). At the completion of the activity called for by that command, the memory controller determines if the background sweep is active. If it is not active, the background status is inspected and action is taken as appropriate, as described later with regard to the background check. Following the background status check, the cache-ahead status is checked, as described later with regard to the cache-ahead management. The controller then waits for the next host command. The controller may not be completely inactive at this time, inasmuch as either the background sweep or the cache-ahead may have initiated or continued some disk activity. If the background was found to be active, its activity is continued until such time as it has no more immediate work to do, as described later with regard to background continuation.

When the background sweep completes a command, the controller is given an interrupt with a signal that indicates the sweep needs its attention. At that time, the controller initiates the next sweep event, if any is waiting, and schedules the next write from cache to disk also based on need, as described later with regard to the background continuation. At the completion of each sweep event, the controller determines if there is a need to continue the sweep. If no such need exists, the background sweep is placed in the dormant state. In either case, when the controller has completed its housekeeping, it becomes inactive awaiting its next task.

The background sweep can be activated in either of two ways; it will be activated when a set number of cached tracks have been modified and are in need of being written from cache to disk. The sweep may also be activated by a timeout. A timeout occurs whenever the sweep is inactive, and there exists any modified track waiting to be written from cache to disk which has been waiting more than a preset amount of time. When a timeout occurs, the controller is signaled that the sweep requires its attention. The controller initiates the background sweep (see description of background initiation) and, after completing the appropriate housekeeping, awaits the next command or event requiring its attention. The background sweep itself continues in operation until there is no more immediate need for its services. At that time it is returned to the dormant state.

Host-Command Management

Whenever a command is received from the host computer, it is given the highest possible priority and handled as depicted in FIG. 5. To determine what actions are required, the command must be analyzed. A portion of the firmware is dedicated to that purpose (see description of host command analysis). The analysis of the command determines the type of command (read, write, seek, or other) and, where meaningful, will make a cache hit/miss determination. The analysis also sets up a table of one or more lines which will be used later in servicing the command.

If the command is a read and it can be serviced entirely from cache (i.e. a cache hit), the command is serviced by the read-hit portion of the controller (see description of read-hit handling).

If any portion of the read cannot be serviced from cached tracks (i.e. a cache miss), the command is turned over to the read- miss portion of the controller (see description of the read-miss handling).

If the command is a write and all tracks involved in the operation are already in cache, the command is serviced by the write-hit portion of the controller (see description of write-hit handling).

If any portion of the write involves an uncached track or tracks, the command is turned over to the write-miss portion of the controller (see description of the write-miss handling).

If the command is a seek, and the target track is already cached, no action is required. If the target track is not cached, the command is turned over to the seek-miss portion of the controller (see description of seek-miss handling).

Analyze Host I/O Command

As depicted in FIG. 6, the analysis of a host command includes creation of a track address list which contains the locations of each track involved in the operation (see description of track address list setup). For each such track, the list contains the track's current location in cache, if it already resides there; or where it will reside in cache after this command and related caching activity have been completed. In the case that a track is not already cached, the space for it to be put into in cache is located, and the current track resident in that space is decached. The analysis includes setting the cache hit/miss flag so that the controller logic can be expedited.

Set Up Track Address List

As shown in FIG. 7, the controller segment which sets up the track address list uses the I/O sector address and size to determine the disk track identifying numbers for each track involved in the I/O operation (see description of address translation). The number of tracks involved is also determined, and for each track, the portion of the track which is involved in the operation is calculated.

Address Translation

FIG. 8 describes the operation for this translation. A sector address can be converted into a track address by dividing it by the track size. The quotient will be the track number, and the remainder will be the offset into the track where the sector resides.

Cache Read-Hit Operation

Refer to FIG. 9. A read hit is satisfied entirely from the cached data. In order to reach this module of the controller, the command will have been analyzed and the track address table will have been set up. With this preliminary work completed, the host read command can be satisfied by using each line of the track address table as a subcommand control. Since all required portions of all affected tracks are already in cache, all required data can be sent directly from the cache to the host. In addition to transferring the data to the host, this module will rechain the affected tracks to become the most-recently-used tracks in the LRU table. Finally, the recycle register value is adjusted according to the recycling rules in effect. As a minimum, in the simplest case the recycle register is set to one to indicate that this cache track is to be considered for recycling. If more complex recycling rules have been specified, the recycle register is incremented according to those rules.

Cache Read-Miss Operation

A cache read-miss (FIG. 10) is satisfied in part or wholly from the disk. In order to reach this module of the controller, the command will have been analyzed and the track address table will have been set up. With this preliminary work completed, the host read command can be satisfied by using each line of the track address table as a subcommand control. For an I/O whose size exceeds the READ-MISS-MAXSIZE, uncached portions are sent directly from the disk to the host without affecting the cache in any way. For an I/O whose size does not exceed the READ-MISS-MAXSIZE, the operation is handled based on the caching device's current mode.

If the unit is not in urgent mode: For track segments which are already in cache, the data can be sent directly from the cache to the host. For a track segment not resident in the cache, the data is sent from the disk to the host, and simultaneously, the portion of the track from the first sector of the requested data to the end of that track is sent to the cache. The LRU-CACHED-LOW and LRU-CACHED-HIGH fields of the corresponding LRU table line(s) are set to reflect the portions of those tracks which have been brought into cache.

If the unit is in urgent mode: For a track not resident in the cache, the data is sent directly from the disk to the host without being entered into the cache.

In either mode, in addition to transferring the data to the host, this module will rechain affected, cached tracks to become the most-recently-used slots in the LRU table.

Cache Write-Hit Operation

A Cache Write-Hit (FIG. 11) is handled entirely within the cache. In order to reach this module of the controller, the command will have been analyzed and the track address table will have been set up. With this preliminary work completed, the host write command can be satisfied by using each line of the track address table as a subcommand control. Since all affected tracks are already represented in cache, all data can be sent directly from the host to the cache without any concern for post-transfer staging of partial tracks. In addition to transferring the data to the cache, this module will, if the slot was linked into the LRU chain, remove the affected cache slot from the LRU chain. In every case, the corresponding LRU-MOD-LOW, LRU-MOD-HIGH, LRU-CACHED-LOW, and LRU-CACHE-HIGH fields are updated to reflect the existence of this new data. Finally, the recycle register value is adjusted according to the recycling rules in effect.

Cache Write-Miss Operation

If the I/O size exceeds the WRITE-MISS-MAXSIZE, uncached tracks or track segments are written directly to disk with no effect on the cache. For an I/O whose size does not exceed WRITE-MISS-MAXSIZE, the operation is handled based on the caching device's current mode. If the caching device is operating in normal mode, a write miss is handled entirely within the cache but requires the placing of information into the LRU-CACHED-LOW, LRU-CACHED-HIGH, LRU-MOD-LOW, and LRU-MOD-HIGH fields to insure that any data required to fill gaps between the modified portions of any partial tracks can be read from the disk into the cache memory, as depicted in FIG. 12. In order to reach this module of the controller, the command will have been analyzed and the track address table will have been set up. With this preliminary work completed, the host write command can be satisfied by using each line of the track address table as a subcommand control. Since this is a cache-miss, some or all of the affected tracks are not in cache; however, all data can be sent directly from the host to the cache.

The cache controller has the responsibility for reading from disk into cache that data required to fill any gaps in the modified portion of tracks. In actuality, only the first and/or last tracks involved in the transfer can be partial tracks; all interior tracks must be full tracks, and thus require no data to be read from the disk to fill gaps in the modified portion of the cached data in any case. For those tracks requiring post-transfer staging, the controller sets up a list of caching events to bring any required track segments into cache to maintain the integrity of the cached tracks. In addition to transferring the data to the cache, this module removes the affected tracks from the LRU chain. If the caching device is operating in urgent mode, the handling of a write miss bypasses the cache for any tracks which are not currently cached, sending the data directly from the host to the disk. The LRU and ADT tables are updated for any cached tracks which may have been affected. Finally, the recycle register value is adjusted according to the recycling rules in effect.

Seek Cache Miss

As shown in FIG. 13, the controller has the option of ignoring a seek command since the controller will ultimately be responsible for fulfilling any subsequent, related I/O command. For a seek command for which the target track is already in cache, no controller action is needed or appropriate. For a seek command for which the target track is not in cache, the controller, if the disk to which the seek is directed is not busy, will cache that target track. This action is based on the assumption that the host would not send a seek command unless it was to be followed by a read or a write command. If the disk to which the seek is directed is busy when a seek command is received from the host, the seek command is ignored.

Recycling of Cached Tracks

It is a well known concept that the basic LRU methodology is a way to keep data in cache based on its very recent history. The recycle feature extends the amount of history considered for data retention with the older portion of history having a lesser effect on the decision to retain the data in cache. The recycling of currently cached tracks is a method of maintaining in cache the tracks of data which have the highest likelihood of being reused.

Recycling Based on Simple Rules

In its simplest form, any cached track which has been read from or written into since it was most recently cached or similarly reused since it was last moved to the MRU position is a candidate for recycling. When a cached track is reused, the recycle register is set and the cached track is moved to the MRU position in the LRU table. When a cached track reaches the LRU position in the LRU table, its recycle register is checked. If the recycle register is zero, indicating the cached track was not reused in its most recent trip from the MRU position to the LRU position, the cached track is a candidate for reuse. If, however, the recycle register is nonzero, indicating the cached track had recently been reused, the cached track is moved to the MRU position of the LRU table, and the recycle register is cleared.

In this manner, a track of data whose recent history indicates reusage is allowed to remain in cache for a longer time than a track of data without such recent use. The effect of this procedure is to allow tracks containing data which has had a low frequency of usage to be reused for caching new material in preference to reusing those tracks whose data have exhibited a higher rate of usage, and to allow the more frequently used data to remain in cache for a longer time.

Recycling Based on Complex Rules

In its more complex form, the recycle register for each cache track is more than a single bit, allowing for a maximum value of n based on the number of bits allocated to the register. The recycle register value is set or adjusted based on certain events in a cached track's life. For example,

1. When a cache track is initially assigned to a specific disk track, the cache track recycle register is given some initial value.

2. When a cached track is rehit, such as by a read or write hit, its recycle register value is adjusted by some amount to increase the tracks chances of being retained in cache at some future time when it would otherwise be a candidate for decaching.

3. When a cached track reaches the LRU position in the LRU table, and if, based on the recycle register value, the cached track is not decached, then, as a penalty for not being rehit during its last trip from the MRU position to the LRU position, its recycle register value is adjusted by some amount to reduce the track's chances of being retained in cache upon some future examination.

The important of the recycle register is in making the correct determination of whether or not to decache a cached track when it has reached the LRU position in the LRU table.

When a cached track reaches the LRU position in the LRU table, its recycle register is examined. If the recycle register value does not meet some set criteria, such failure indicating the history of the cached track does not justify retention in cache, the cached track is a candidate for decaching and its space can be made available for reuse. If, however, the register value meets the preset condition, indicating the cached track has had a level of activity that indicates the cached track should be retained in cache for some more time, the cached track is moved to the MRU position of the LRU table, and the recycle register value is adjusted by some amount to indicate the track was moved to the MRU position based on its history, rather than due to a reusage such as an instant read hit.

Some of the factors to be considered in determining the initial setting of the recycle register when the track is first cached are:

1. The reason for caching the track such as read miss, write miss, or read-ahead;

2. The current system status, such as the proportion of cache which is now in a modified condition and needs written to the disk.

Some of the factors to be considered in determining the amount of adjustment of the recycle register value when the cached track is reused are:

1. The current value in the recycle register;

2. The original reason for caching the track such as read miss, write miss, or read-ahead;

3. The nature of the most recent activity of the cache track, such as a read hit, it has just been written to disk following one or more writes from the host into it, or that it has just been cached by a read-ahead operation.

4. The current system status, such as the proportion of cache which is now in a modified condition and needs written to the disk.

Some of the factors to be considered in determining the action to be taken, and the amount of adjustment to the recycle register value if any, when the cached track reaches the LRU position are:

1. The current value in the recycle register;

2. The original reason for caching the track such as read miss, write miss, or read-ahead;

4. The current system status, such as the proportion of cache which is now in a modified condition and needs written to the disk.

Example of Usage of the Recycle Register

The following is one possible example of a set of the controlling factors and corresponding register adjustments for one incarnation of the recycle register concept. This is a simple example which favors quick release of data cached by read-ahead, and also reduces retention of cached material when the caching device is encountering heavy data modification (host writes). For the sake of illustration in this example, one can assume P% to be 50%.

1. When a cache track is initially assigned to a specific disk track, the recycle register is set according to the following table:

2. Adjustment to recycle register when rehit (in all cases of a rehit, the track is moved to the MRU position):

3. Adjustment to recycle register when track reaches LRU:

Decache a Track

For every cache-miss I/O that occurs, and for every cache-ahead operation, some previously cached track or tracks of data must be decached. The primary function of the LRU table is to allow the controller to expeditiously determine which cached track of data is the best candidate for decaching. The decaching module depicted in FIG. 14 chooses the track to be decached. Normally, the track with the longest elapsed time since its last usage will be the track to be decached. This is the track which is pointed to by the LRU-LRU element of the LRU table. The LRU-LRU pointer is maintained in such a way as to always point to the least-recently-used track. In the described device, the decision of which track to decache is based not only on the LRU concept, but also takes into account the recycling factors as described in the sections of this document regarding recycling.

The first condition that must be met in order to decache a track is that the track must be inactive and the cached data must be unmodified; that is, it is not at this instant the subject of any activity such as being written from the cache to the disk. It is highly unlikely that the LRU-LRU track would be the current subject of any activity since most activities on a track reposition it out of the LRU-LRU slot. Also, for modified data, it is most likely it would have been written to disk before the cached track has moved down through the LRU table to the LRU position. However, the possibility is covered by the decaching algorithm.

The second condition that must be satisfied before a track can be decached is that the track to be decached must not be a candidate for recycling. The recycling mechanism provides for moving cached tracks from the LRU position to the MRU position in the LRU table. Normally, recycling will be done as a background task. In the event a track which should be recycled is found in the LRU position when a track is needed for caching new data, that recycle candidate can be recycled at that time. The recycling mechanism described above will determine the desirability of recycling of each track as it reaches the LRU position. When a track reaches the LRU position, the recycle mechanism also makes the appropriate adjustment to the recycle register. As stated in the recycling description, the effect of this procedure is to allow the tracks containing the more frequently used data to remain in cache for a longer time.

If, for any of the above reasons, the LRU-LRU track is unable to be decached, the LRU chain will point to the next-to-LRU track. While it is possible for the LRU track to be a candidate for recycling, it will be an unusual situation in which the decaching module will need to inspect more than one LRU slot to find the track to be decached.

When the actual candidate for decaching has been identified, both the LRU and ADT tables are updated to reflect that the chosen candidate is no longer cached. This is a minor amount of work; no disk activity is involved.

Cache-Ahead Management

The cache hit and management operation is depicted in FIG. 15. The controller attempts to cache-ahead after every host I/O which is a read operation regardless of whether the I/O was a cache hit or a cache miss. Operations which write data from the host to the device need no cache-ahead operations since data can always be accepted from the host into the cache's SSD. However, a read cache-ahead action is a background type of activity, and only uses the private channel between disk and cache, it will have a very minimal negative impact on the caching device's response time to host I/O activity. To further limit the impact, the cache-ahead is given a lower priority than any incoming host I/O request.

A major factor in limiting the cache-ahead activity is the lack of need for its operation following most host I/O's. As depicted in FIG. 16, the caching device determines the number of data segments of the same size as the current host I/O which remain between the location of the end of the current host I/O data and the end of the cached track containing that data. If this computed number of data segments is more than a predetermined number, the cache unit can handle that number of host I/O's before there is a need to fetch data for the succeeding track from the disk into the cache memory.

If, on the other hand, the computed number of data segments is not more than the predetermined number, it is possible for the host to access all those segments between the end of the current host I/O data location and the end of the cached track in the same or less time than it would take for the caching device to fetch the succeeding track of data from the disk into the cache memory. In this case, the caching device should immediately initiate action to fetch the succeeding data track from the disk so that the service to the host can proceed with the least disk-imposed delays.

Conversely, if the caching device were to ignore the above-described locality factor and always fetch the next data track after every cache read-miss, many unneeded tracks of data would be fetched from disk into cache memory. Such excessive fetches would use up more of the caching device's resources with a negative impact on the caching device's host service time.

There are only two candidates for cache-ahead: they are the single track immediately following that involved in the host I/O and the track immediately preceding that of the host I/O. Since these tracks will often have already been cached by previous cache-ahead activity, the cache-ahead activity is largely a self-limiting process.

Only one track is cached-ahead for any given host I/O: the track succeeding the host I/O is the primary candidate. If it is not already cached, and the proximity factor indicates the cache-ahead should occur, the forward track is cached at this time. If the succeeding track is already cached, the track preceding the host I/O is considered; if it is not already cached, and the proximity factor favors caching, this preceding track is cached at this time. Of course, if both of these candidate tracks had been cached previously, the cache-ahead module has no need to do any caching.

A very important benefit accrues from this cache-ahead, cache-back feature. If related tracks are going to be accessed by the host in a sequential mode, that sequence will be either in a forward or backward direction from the first one accessed in a given disk area. By the nature of the cache-ahead algorithm, an unproductive cache-ahead will only involve one track which lies in the wrong direction from the initial track in any given track cluster. This, coupled with the proximity algorithm, makes the cache-ahead behavior self-adapting to the direction of the accesses.

Background Sweep Management

When a write I/O from the host is serviced by the controller, the data from the host is placed in the cache. It is written from the cache to the disk in the background, minimizing the impact of the disk operations on the time required to service the I/O. The module that handles this background activity is the background sweep module. To limit the sweep activity, and thus limit contention for the spindle, only those portions of tracks which have been modified are written from SSD to disk during a sweep. In the interest of further efficiency, the background sweep module does not always copy data from cache to disk as soon as it is available. Rather, it remains dormant until some minimum number of modified tracks are waiting to be copied before going into action. In order to avoid having a single modified track wait an inordinately long time before being copied from cache to disk, the background sweep will also be activated by a timeout. Thus, if any modified track has been waiting a certain minimum time, and the sweep is not active, the sweep will be activated. After the sweep has copied all modified portions of tracks from cache to disk, it returns to a dormant state.

Sweep Timeout

A timeout occurs when some cached data track has been modified and the corresponding track on disk has not been updated after a certain minimum time has elapsed. When a timeout occurs, by definition there will be at least one cached track which needs to be copied to disk. At this time, the background will be changed into the active state. The timeout module (FIG. 17) also causes the first background event to be set up (see description of background event generation), and if no conflict exists with the host for access to the disk, the execution of the event will be initiated. After this event is initiated, the next event, if one is known to be needed, is also set up and held for later execution. When these things have been done, the background sweep waits for circumstances to cause it to continue its operation or to return to a dormant state.

Sweep Initiation

At the completion of each host I/O operation, the sweep initiation module (FIG. 18) is entered. One of three cases may exist. The first case is that the sweep is dormant, and there are not a sufficient number of modified tracks waiting to be copied to disk to cause the sweep to be enabled at this time. In this case, which is the most common one, there is no action to be taken at this time.

In the second case, the sweep is active, and a background event is operating. In this situation, no new action is needed at this time.

In the final case, the sweep is active, but no background event is currently in operation. Under these conditions, a background event is generated (see description of Generate Sweep Event) and, if appropriate, its execution is initiated.

Generate Sweep Event

The need for the generation of a background sweep event is predicated on there being no other ongoing activity involving the disk. If the event generation module of FIG. 19 is entered when any such activity is in progress, no event is generated.

At times, the event generation module will find that there are no more modified tracks waiting to be copied to the disk. In this case, the background sweep is returned to the dormant condition. At other times, the background sweep is in the active mode, but has been temporarily interrupted to handle the higher priority activity of servicing a host I/O. Such interruption requires the background sweep to be restarted. It does this by finding the modified track which is nearest, but not directly at, the disk head; initiating a seek to that track; and then setting up a write event for the track. This write event will not be initiated until later, but its existence signals the sweep continuation module (see description of continuation module) that, if possible, this write is the next thing to be done.

The effect of this method of handling background writes is to minimize the impact on the host operations. The controller has an opportunity to service host I/O misses between the background seek and the corresponding write operation. None of this has any significant effect on servicing host I/O cache hits since hits are always handled immediately. The disk is not involved in a hit.

The sweep handles the writing of modified tracks differently depending on whether all the sectors in the track have been modified, or only some of the sectors have been modified. Further, the number of wholly modified tracks and the number of partially modified tracks are both taken into consideration in the setting of priorities for writing individual tracks to disk. When a larger number of wholly modified tracks exist, as opposed to the number partially modified, the wholly modified tracks are given preference by the sweep operation.

Writing a modified track from cache to disk is limited to handling only the modified portion of the track as defined by the corresponding LRU-MOD-LOW and LRU-MOD-HIGH values. Once the modified track or track segment has been written to disk, the track's cache slot, which has been in an unchained status, is placed in the LRU chain at the MRU position if the track had been only partially modified, and is placed at the LRU position if the track had been wholly modified. At the same time, the corresponding LRU-MOD-LOW and LRU-MOD-HIGH fields are set to their null value to indicate that no part of the cached data differs from that in the corresponding disk track.

Sweep Continuation

As depicted in the flow chart of FIG. 20, each background sweep event, whether a seek or a write, prepares a waiting event for the sweep's subsequent action. Thus, the initiation of a seek always prepares the subsequent, related write event; the initiation of a write prepares the subsequent, unrelated seek event, if another track is waiting to be copied to disk.

The continuation module is entered upon the completion of each sweep event. If the host has issued an I/O command which requires the disk (in other words, a cache-miss), the background sweep sequence is interrupted, and the waiting event is erased. This action is taken in order to expedite the servicing of the host's commands, and is taken regardless of the type of sweep event which is waiting. It can result in wasting background seek actions. This is acceptable; the aborted write will be handled later when time permits. 0f course, once a sweep command, whether a seek or a write, has actually been initiated, it cannot be aborted.

If the sweep continuation module is entered after the sweep operations have been interrupted, it will use the event generation module (see description of event generation) to restart the sweep sequence.

Finally, if the continuation module finds that the just completed sweep operation was a write, and no more modified tracks are waiting to be copied to the disk, the sweep is put into the dormant state.

Power Down Control

As depicted in the flow chart of FIG. 21, this portion of the firmware is invoked when the unit senses that the line power to it has dropped. Since some of the data in the unit may be in the cache portion in a modified state and awaiting transfer to the disk, power must be maintained to the cache memory until the modified portions have been written to the disk. Thus, a failure of the line power causes the unit to switch to the battery backup unit. The battery backup unit provides power while the memory device goes through an intelligent shutdown process.

If the host is in the process of a data transfer with the memory device when power drops, the shutdown controller allows the transfer in progress to be completed. It then blocks any further transactions with the host from being initiated.

The shutdown controller then must initiate a background sweep to copy any modified portions of data tracks from the solid state memory to the disk so that it will not be lost when power is completely shut off to the control and memory circuits. After the sweep is completed (which will take only a few seconds), all data in the solid state memory will also reside on the disk. At this point the disk spindle can be powered down, reducing the load on the battery.

Most power outages are of a short duration. Therefore, the controller continues to supply battery power to the control circuits and the solid state memory for some number of seconds. If the outside power is restored in this time period, the controller will power the spindle back up and switch back to outside power. In this case, the operation can proceed without having to reestablish the historical data in the solid state memory. In any case, no data is at risk since it is all stored on the rotating magnetic disk before final shutdown.

Final Background Sweep

The final background sweep (FIG. 22) copies modified portions of tracks from the solid state memory to the magnetic disk. There will usually be only a few such tracks, or portions of tracks to copy since the number that can reach this state is intentionally limited by the operations of the system. The final sweep makes use of logic developed for the normal operation of the background sweep.

The sweep is initiated in much the same manner as for a timeout during normal operation. If no tracks need to be copied, the sweep is left in the dormant state, and no further sweep action is required. If any tracks need copied, the sweep initiator sets up and initiates the first background seek, as well as sets up the related write event. At the completion of this first seek, control goes to the background continuation module which alternately executes the previously created, waiting event and generates the next event and puts it into a wait status. When no modified tracks remain to be copied, the sweep is finished.

Parameters and Particulars

This specification refers to items which are not given specific quantities or identities. These have been purposely left unquantified so as not to imply any absolute limits or restrictions. For purposes of illustration, and to provide known workable dimensions and identities, the following ranges of values and identifiers are provided, along with a set which is satisfactory for a sample configuration.

Background Sweep Trigger, Number of Modified Tracks

Range: One to number of tracks on chosen disk.

Sample configuration: Five

Background Sweep Trigger, Time

Range: One millisecond to unlimited.

Sample configuration: Five seconds.

Eprom Memory for Microprocessor

Size range: Non-specific.

Sample configuration: 64 kilobytes.

Hardware Microprocessor Controller

Candidates: Any suitable and available microprocessor.

Sample configuration: 80C196, 24 Mhz (Intel Corporation of Santa Clara, Calif.).

Power Down, Cache Hold Time

Range: Zero seconds to limitation imposed by battery backup unit.

Sample configuration: Five minutes.

Rotating Magnetic Disk Capacity

Size range: Any available disk capacity.

Sample configuration: 675 megabytes formatted.

Scsi Controller

Candidates: Any suitable and available controller device.

Sample configuration: NCR 53C90A (National Cash Register Corporation, Dayton, Ohio).

Sector Size

Size range: Any appropriate for the host system and the selected disk drive.

Sample configuration: 180 bytes.

Sectors Per Track

Range: Any appropriate for selected disk and host system.

Sample configuration: 256.

Solid State Memory Size

Size range: One megabyte to 100 percent of the capacity of the attached disk capacity.

Sample configuration: 32 megabytes.

Track Size

Size range: One sector to any size appropriate for the selected disk drive.

Sample configuration: 256 sectors.

Tracks Per Disk

Range: Any available on chosen disk.

Sample configuration: 14628.

Table Formats

The invention now being fully described, it will be apparent to one of ordinary skill in the art that many changes and modifications can be made thereto without departing from the spirit or scope of the appended claims.

Claims

What is claimed:

1. A method for operating a cache memory system having a mass storage device and a cache memory, said method comprising the steps of: storing in said cache memory selected data from said mass storage device; maintaining a Least Recently Used (LRU) table indicating the relative recency of use of each data entry stored in said cache memory; updating a value in a recycle register for each entry in said LRU table upon the occurrence of one or more of: when data stored in said cache memory associated with said LRU entry has been accessed; when data associated with said LRU entry is added to said cache memory; when said LRU table entry is rechained for reuse after a previous non-use; and when said LRU table entry reaches the LRU position of said LRU table; and when cache space is determined to be needed corresponding to an area of mass storage which is not currently assigned in said cache memory: examining the entry at the LRU position of said LRU table to determine the value of said recycle register associated with said LRU table entry; if said value is greater than a predefined threshold value, updating said value and placing said entry at the top of said LRU table and repeating said step of examining the entry at the LRU position of said LRU table to determine the value of said recycle register; and if said value is not greater than said predefined threshold value, decaching said data in said cache memory associated with said LRU table entry, reusing said LRU entry for satisfying said current need for storing data not stored in said cache memory, and placing the new LRU entry at the MRU position of said LRU table or marking it dechained depending on the new usage of the entry.

2. A method as in claim 1 which further comprises the step of maintaining data in said LRU table indicating which sectors of each block of data stored in said cache have been modified.

3. A method as in claim 1 wherein, for each block stored in said cache memory, not all sectors of a given block need be cached.

4. A method as in claim 1 wherein an LRU track entry is flagged as a dechained member of the LRU table if that track contains modified data.

5. A method as in claim 4 wherein a track entry is made a chained member of the LRU table after its modified data is written to said mass storage device.

6. A method as in claim 5 wherein said track entry is made a chained member at the bottom of the LRU table if every sector of said track was modified prior to writing to said mass storage device.

7. A method as in claim 5 wherein said track entry is made a chained member at the top of the LRU table if less than every sector of said track was modified prior to writing to said mass storage device.

8. A method as in claim 1 wherein said step of updating a value in a recycle register upon data associated with said LRU entry being added to said cache memory is based upon the original reason for caching said data within said cache memory.

9. A method as in claim 8 wherein said reason is selected from the group of reasons consisting of a read-miss, a write-miss, a read-ahead, and a read-behind.

10. A method as in claim 1 wherein said step of updating a value in a recycle register upon data associated with said LRU entry being added to said cache memory is based upon the proportion of cache memory which is in a modified condition and needs to be written to said mass storage device.

11. A method as in claim 8 wherein said step of updating a value in a recycle register when said LRU table entry is rechained for reuse after a previous nonuse is dependent on one or more of: the current value in said recycle register; the original reason for caching data associated with said recycle register; the nature of the most recent activity of said data associated with said recycle register; and the proportion of cache memory which is in a modified condition and needs to be written to said mass storage device.

12. A method as in claim 11 wherein said original reason for caching said data is selected from the group of reasons consisting of read-miss, write-miss, read-ahead, and read-behind.

13. A method as in claim 11 wherein said nature of the most recent activity of said data associated with said recycle register is selected from read-hit, write-hit, read-ahead, and read-behind.

14. A method as in claim 11 wherein said original reason for caching said data is selected from the group of reasons consisting of read-miss, write-miss, read-ahead, and read-behind.

15. A method as in claim 11 wherein said nature of the most recent activity of said data associated with said recycle register is selected from read-hit, write-hit, read-ahead, and read-behind.

16. A method as in claim 1 wherein said step of updating a value in a recycle register when said LRU table entry reaches the LRU position is dependent on one or more of: the current value in said recycle register; the previous reason for caching data associated with said recycle register; the nature of the most recent activity of said data associated with said recycle register; and the proportion of cache memory which is in a modified condition and needs to be written to said mass storage device.

Patent Citations (29)

Patent	Date	Inventor
US4086629	1978-04-01	Desyllas
US4394733	1983-07-01	Swenson
US4433734	1984-02-01	Hanson et al.
US4437155	1984-03-01	Sawyer et al.
US4458307	1984-07-01	McAnlis
US4468730	1984-08-01	Dodd
US4490782	1984-12-01	Dixon
US4523206	1985-06-01	Sasscer
US4530054	1985-07-01	Hamstra et al.
US4583166	1986-04-01	Hartung
US4603380	1986-07-01	Easton
US4607346	1986-08-01	Hill
US4611289	1986-09-01	Coppola
US4636946	1987-01-01	Hartung et al.
US4835686	1989-05-01	Furuya et al.
US4843542	1989-06-01	Dashiell
US4882642	1989-11-01	Tayler et al.
US4888691	1989-12-01	George et al.
US4908793	1990-03-01	Yamagata
US4920478	1990-04-01	Furuya et al.
US4956803	1990-09-01	Tayler et al.
US4959774	1990-09-01	Davis
US4972364	1990-11-01	Barrett et al.
US5043885	1991-08-01	Robinson
US5133060	1992-07-01	Weber
US5224217	1993-01-01	Zangenehpour
US5309451	1994-05-01	Noya et al.
US5410653	1995-04-01	Macon, Jr. et al.
EPX66766	1981-05-01

Non-Patent Literature (18)

International Business Machines Corp.--"A Multiple Workload Approach to Cache Planning" by B. McNutt & J. Murray, pp. 9-15, Dec. 1987.
CMG '87--The Computer Measurement Group, Inc. Orlando Dec. 7-11, 1987--BEST/1 Analysis of the IBM 3380-13 Cached Storage Controller by J. Buzen & Anneliese Von Mayrhauser, pp. 156-173.
Computer Science Division, University of California, Berkeley, "Maximizing Performance in Striped Disk Array" by P. Chen & D. Patterson, pp. 322-331, 1990.
University of California, Berkeley, "Cache Memory Design: An Evolving Art", by J. Smith, pp. 40-44, 1987.
University of Wisconsin, "A Case for Direct-Mapped Caches" by M. D. Hill, Dec. 1988, pp. 25-40.
IBM, Almaden Research Center and Tucson Lab., "The IBM 3990 Disk Cache" by Jai Menon and Mike Hartung, pp. 146-151, 1988.
The Computer Society of IEEE, Feb. 29-Mar. 4, 1988 "COMPCON88", Nippon Electric Co., Tokyo, Japan, Integrated Disk Cache System with File Adaptive Control by Tokunaga et al., pp. 412-416.
1988 IEEE, Amperif Corp., "The Amperif Cache Disk System", by S. Fuld, pp. 156-157.
COMPCON88, Feb. 29-Mar. 4, 1988, Boole & Babbage, Inc. "DASD Cache Controllers: Performance Expectations and Measurements" by Kovach, pp. 74-78.
Chen et al., "Maximizing Performance in a Striped Disk Array", 1990, pp. 322-331.
Buzen et al., "Best/1 Analysis of The IBM 3880-13 Cached Storaged Controller", 1987, pp. 156-173.
McNutt, et al., "A Multiple Workload Approach to Cache Planning", pp. 9-15.
CMG '86 Las Vegas Dec. 9-12 1986--Degradation of IMS Performance Using a 3880-23Cache Controller by Carolyn Hanna, pp. 15-21.
CMG '86 Las Vegas Dec. 9-12 '86--Performance & Tuning of Cached I/O Subsystems by M. B. Friedman, pp. 717-727.
CMG XV International Conf. of San Francisco Dec. 1984--Method for the Deployment of IBM 3880 Model 13 Cached Storage Controllers by P. T. Guttman, pp. 44-53.
CMG Transactions Fifty-four, Fall 1986--Disc Caching in the System processing Units of the JP 3000 Family of Computers by John R. Busch et al., pp. 21-39.
Research Report: RC 10197 (#45223) Sep. 21, 1983 ComputerScience--the Access Time Myth, by R. A. Scranton & D. A. Thompson, pp. 1-8.
IBM Systems Journal vol. 24, Nos. 3/4 1985 "Cache-DASD Storage Design for Improving System Performance", Grossman, pp. 316-334.

Abstract

Description