Optimized allocation of data elements among cache lines

[0002] This application is related to the application entitled “Optimized Logging Of Data Elements To A Data Storage Device,” U.S. Ser. No. 09/024,248, by Michael Andrew Brian Parkes, Barry Michael Nolte and Douglas Stewart Boa, assigned to Microsoft Corporation, and filed concurrently with this application. The disclosure of this application is hereby incorporated by reference.

[0003] This invention relates to computer programming and more particularly, determining an optimized assignment of data elements within structures, unions and classes.

[0004] Users are demanding increased performance of their applications running on their computers. Computer hardware, including central processing units (CPUs), are becoming increasingly faster. However, their performance is limited by the speed at which data is available to be processed. In a typical computer, Level 1 (L1) and Level 2 (L2) cache memories are physically close to a processor to provide data at very high rate. The cache memory is typically divided into 32 byte cache lines, a cache line being the common unit of data retrieved from memory. When the required data is not available in L1 cache, a cache line fault occurs and the data must be loaded from lower speed L2 cache memory, or relatively slow RAM. The application is often effectively stalled during the loading of this data, and until such time as the data is available to the CPU. Therefore, by decreasing the number of cache faults, an application will run faster. Furthermore, data elements within an application are not randomly accessed. Rather, data elements, especially within the same structure, union, or class are typically accessed within a short period of other data elements within the same structure, union, or class.

[0005] The first step in optimizing an application is to model the usage patterns of data elements by the application. To accomplish this, the application being optimized is executed and used in a typical manner, with data being recorded that tracks the order in which the data elements are accessed. In doing so, a stream of data at a rate of 30-40 megabytes per second on typical hardware is generated. Traditional disk writing methods cannot keep up with this volume of data. Hence, if all of this data is to be collected to disk, either the disk logging process must also be optimized, or the execution of the application must be slowed down or modified which degrades the accuracy of the data usage model. Therefore, the preferred approach is to optimize the data logging such that the application being modeled is not hindered by the data logging method used.

[0006] Traditional data logging methods operate in a linear fashion by generating a first record of data, writing the first record of data to disk, generating a second record of data, writing the second record of data to disk, and so on. While this approach is simplistic, this method does not optimize the writing of a large amount of data to disk, such as the voluminous data stream generated when modeling the application. In fact, the processing overhead is so high that the linear data writing approach does not allow data to be written at the fastest rate allowed by the hardware. Rather, the data logging rate is limited by software processing of individual write operations. Such is the same problem with reading in records of data, one record at a time. Needed is a solution for writing and reading of a large amount of linear, order dependent data at high rates of speed which approach the physical limitations of the hardware device to which the data is being logged.

[0007] According to the invention, cache line groupings of data elements within a single structure, union, or class of an application are determined for minimizing the frequency of cache line faults. As used herein, the meaning of the terms “structures” or “structure” include structures, unions, classes, or the like. Furthermore, this disclosure describes the invention in terms of optimizing the performance of a computer “application”. However, the invention disclosed herein is equally applicable to any computer program including end-user applications, operating systems, objects, drivers, operating environments, library routines or objects, and the like.

[0008] Data is first collected describing the application's usage of data elements within each structure. Next, this data is manipulated to determine statistical correlations among accesses to data elements within the same structure. Optimized cache line groupings of data elements are then determined in order to maximize the probability that for any data element accessed within a structure, the next or previously accessed element will be within the same cache line. Using this optimized grouping, the source code of the application is edited to re-order the declaration statements such that the re-built application will have optimized groups of data elements assigned to cache lines. In this manner, the application will generate fewer cache line faults, and run more efficiently.

[0009] More specifically, an application is executed and used in a manner characteristic of a typical use of the application with accesses to the data elements being recorded. To generate this data, the source code of the application is first compiled with instrumentation being added so that a data record is produced each time a data element is accessed. Then, when the application is subsequently executed, the application generates a disk file containing a sequential stream of data records containing an entry for each time a data element is accessed and the type of access (a read or write operation). Thus, two data elements accessed one after the other will have corresponding sequential entries within the stream of data. Alternatively, a background process could be used to track the accessing of data elements.

[0010] To keep up with the 30-40 megabytes of data produced per second on typical current hardware, an optimal data logging process is used to efficiently write the stream of data to disk at a rate which approximates the maximum rate that the disk hardware can support, which cannot be achieved using traditional data writing methods. To overcome the processing limitations on the transfer rate of traditional data logging methods, the performance of a write operation is divided between data source and data logger software processes which operate in different threads, these threads being in the same process or in different processes. The data source first retrieves a buffer from a pool of empty buffers; each buffer containing references to blocks of contiguous memory addresses.

[0011] Further efficiency is gained when the size of each block of memory addresses corresponds the file allocation size (or a multiple thereof) for the hardware device (e.g., a disk) being employed. Normally, when a file is expanded on a hardware device, the additional space is allocated and the contents of this addition space on the disk must be erased for security reasons (i.e., to protect the prior user data stored in this newly allocated space). However, by combining (1) a file allocation request and (2) a data storage request for writing to the entire contents of the additional space requested into a single operation, the data erasing step is unnecessary and can be eliminated to increase data logging efficiency. In an embodiment of the present invention, each block of memory addresses corresponds to a page of memory which also corresponds to a file allocation size for the hardware device. A typical block of memory addresses used in an embodiment of the present invention is four (4) Kbytes in size, and will vary depending on the hardware platform on which the present invention is being practiced.

[0012] The data source then fills this memory block with data records, and when full, the buffer is placed at the end of a queue of full buffers to be written to the hardware device.

[0013] When there are buffers in the queue of full buffers, the data logger, operating a synchronously with respect to the data source, consolidates (i.e., packages) the full buffers into larger data blocks. In some instances, these larger data blocks will reference non-contiguous blocks of memory addresses (i.e. the memory addresses corresponding to the buffers comprise a non-contiguous address space). In one embodiment, an array is filled with pointers to the blocks of memory addresses contained in the buffers which were removed from the full buffer queue. For simplicity and efficiency reasons, each larger data block corresponds to the same fixed size (e.g., 64 Kbytes or 16 pages of memory). The size of the consolidated blocks is adjustable, and is preferably tailored to match the characteristics of the hardware device and its controller to achieve maximum performance.

[0014] A described herein, efficiency is gained by logging chunks of data in sizes equivalent to the file allocation size (or multiples thereof) of the hardware device.

[0015] The data logger then logs the corresponding data records to the hardware device by passing the consolidated data block to a hardware controller to efficiently store the data on the hardware device. This logging of the data is performed using direct memory accesses (DMA). Such a DMA memory accesses do not involve the computer CPU, and allow data transfer directly between memory and a peripheral device, such as a disk drive. In the Windows NT environment, the WriteFileGather operation, described in detail hereinafter, is used to synchronously write the data to disk with the consolidated block corresponding to an array of memory page pointers and a count of the number of array entries.

[0016] By decoupling the data source from the physical write operations to the disk, the data source is able to use a series of asynchronous data store operations to empty buffers, while the data logger retrieves large buffers from the full buffer queue and synchronously (or even a synchronously) writes the large buffers without stalling the data source. Using this method, the hardware data logging device is able to perform optimally up to the limits of the physical characteristics of the hardware logging device while creating or expanding a data file. Various embodiments in keeping with the invention disclosed herein include single or multiple data source and data logging processes with single or multiple empty buffer sets and full buffer queues.

[0017] In one embodiment of the invention, a total of 64 buffers (corresponding to 64 pages of memory) comprise the empty buffer pool and the full buffer queue. In keeping with the invention, the total number of buffers (and memory pages) is adjustable, with this number preferably sized to match the desired performance and memory requirements. Such matching is a classical allocated memory versus performance trade-off. If there is enough available memory given the memory requirements of the application itself, enough buffers and memory pages should be allocated such that the application is never stalled waiting for an empty buffer to which to store its data.

[0018] In keeping with the scope and spirit of the invention disclosed herein, the optimal data logging architecture and methods can also be used to efficiently read blocks of data from a hardware device. In this case, the ReadFileScatter operation (described in detail hereinafter) is used in place of the WriteFileGather operation, and the data logger receives the stream of data from the full buffer queue, with the data logger filling empty buffers with the read data.

[0019] Once the data characterizing the accesses of data elements has been collected, the recorded data stream is then manipulated to generate a data structure containing statistics describing the relationships among accesses to the structures and/or instances of structures with respect to the timing and frequency of their being accessed. In an embodiment, the data structure is built containing statistics describing the number of times pairings of intra-structure data elements are accessed within some predefined number of CPU data operations, and whether a read or write operation was performed.

[0020] For example, if there are two structures declared in a program such as:

[0021] struct example1 {int A,B,C,D,E,F,G,H,I,J,K,L;} x, y;

[0022] struct example2 {int A,B,C,D,E,F,G,H,I,J,K,L,M;} z; the invention described herein will determine an optimized ordering of integers A-L within “example1”, and A-M within “example2”. In determining the optimized ordering, the attributes of the accessing of elements from within a single instantiation of a structure are analyzed. If x.A is usually accessed immediately before or after x.F, then F and A are preferably defined within example1 such they are assigned to the same cache line.

[0023] In determining the optimized groupings of intra-structure data elements, a series of linear equations are derived which provide a correlation among data elements within the same structures with respect to their accessing attributes. To derive these equations, the data stream is processed as follows.

[0024] Next, for each structure, a series of linear equations is derived by first considering whether read or write operations predominate an element's relationship with the other data elements within the structure. By reducing the inclusion of highly written data elements within cache lines with highly read data elements, the number of cache line faults can be reduced. Therefore, linear equations are derived with highly read data elements grouped together and predominately written elements grouped together.

[0025] A data element pairing is defined to be “highly written” if the total number of write operations for the pairing is greater than one-third of the total read operations. Although other ratios can be used, this one was selected because a write operation typically takes twice the amount of time as does a read operation. For example, in a multiprocessor environment, a cache line can simultaneously be in cache memory of multiple CPUs as long as read operations are performed on the data elements within the cache line. However, a CPU must have exclusive access to a cache line before it can write to it. These linear equations are then solved based using the number of read and write operations to achieve optimized cache line groupings of highly read and highly written intra-structure data elements. These results are then used to produce recommended declaration statements for each structure. After manual or automatic editing of the application's source code, the application is re-compiled and linked. Alternatively, the optimized grouping could be used as input to a compiler to automatically produce optimized cache line groupings.

[0026] The appended claims set forth the features of the present invention with particularity. The invention, together with its advantages, may be best understood from the following detailed description taken in conjunction with the accompanying drawings of which:

[0027] FIGS. 1, 2A and 2B are block diagrams of an exemplary operating environments in which the invention may be practiced, including a computer system for optimally logging data, determining an optimized allocation of data structure elements within cache lines, and running an optimized application;

[0028] FIG. 3 is a high-level flow diagram describing the steps for determining an optimized allocation of structures within the cache lines of a computer system;

[0029] FIG. 4 is a block diagram of a data record generated by the application to be optimized which describes an operation on a data element of a structure, union, or class;

[0030] FIG. 5A is a block diagram illustrating the basic flow of data between a data source and a data logger;

[0031] FIGS. 5B-5C are block diagrams describing an embodiment of the empty buffer pool and full buffer queue;

[0032] FIG. 6A is a high-level flow diagram describing the processing by the data source;

[0033] FIG. 6B is a high-level flow diagram describing the processing by the data logger;

[0034] FIGS. 7A and 7B describe the WriteFileGather and ReadFileScatter Windows NT commands;

[0035] FIG. 8 is a block diagram of a hash table and array used in processing the data collected from the application to be optimized;

[0036] FIG. 9 is a flow diagram describing the steps for processing the collected data to determine an optimized grouping of data elements within a cache line;

[0037] FIG. 10 contains representative pseudo code for populating the hash table and array for rapid processing of the collected data in accordance with the invention;

[0038] FIG. 11 contains representative pseudo code for creating cache line equations; and

[0039] FIG. 12 is a block diagram illustrating the process of determining how to allocate data elements to cache lines for decreasing the frequency of cache line faults.

[0040] FIGS. 1, 2A and 2B and the following discussion are intended to provide a brief, general description of a suitable computing environment in which the invention may be implemented. Although not required, the invention will be described in the general context of computer-executable instructions, such as program modules, being executed by a personal computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Moreover, those skilled in the art will appreciate that the invention may be practiced with other computer system configurations, including hand-held devices, multiprocessor systems, microprocessor-based or programmable consumer electronics, network PCs, minicomputers, mainframe computers, and the like. The invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote memory storage devices.

[0041] With reference to FIG. 1, an exemplary system for implementing the invention includes a general purpose computing device in the form of a conventional personal computer 20, including a processing unit 21, a system memory 22, and a system bus 23 that couples various system components including the system memory to the processing unit 21. The system bus 23 may be any of several types of bus structures including a memory bus or memory controller, a peripheral bus, and a local bus using any of a variety of bus architectures. The system memory includes read only memory (ROM) 24 and random access memory (RAM) 25. A basic input/output system 26 (BIOS) containing the basic routines that helps to transfer information between elements within the personal computer 20, such as during start-up, is stored in ROM 24. In one embodiment of the present invention, the commands to determine the optimized cache groupings of intra-structure data elements and to perform the optimized logging of data are stored in system memory 22 and are executed by processing unit 21. The personal computer 20 further includes a hard disk drive 27 for reading from and writing to a hard disk, not shown, a magnetic disk drive 28 for reading from or writing to a removable magnetic disk 29, and an optical disk drive 30 for reading from or writing to a removable optical disk 31 such as a CD ROM or other optical media. The hard disk drive 27, magnetic disk drive 28, and optical disk drive 30 are connected to the system bus 23 by a hard disk drive interface 32, a magnetic disk drive interface 33, and an optical drive interface 34, respectively. The drives and their associated computer-readable media provide nonvolatile storage of computer readable instructions, data structures, program modules and other data for the personal computer 20. Although the exemplary environment described herein employs a hard disk, a removable magnetic disk 29 and a removable optical disk 31, it should be appreciated by those skilled in the art that other types of computer readable media which can store data that is accessible by a computer, such as magnetic cassettes, flash memory cards, digital video disks, Bernoulli cartridges, random access memories (RAMs), read only memories (ROM), and the like, may also be used in the exemplary operating environment.

[0042] A number of program modules may be stored on the hard disk, magnetic disk 29, optical disk 31, ROM 24 or RAM 25, including an operating system 35, one or more application programs 36, other program modules 37, and program data 38. A user may enter commands and information into the personal computer 20 through input devices such as a keyboard 40 and pointing device 42. Other input devices (not shown) may include a microphone, joystick, game pad, satellite dish, scanner, or the like. These and other input devices are often connected to the processing unit 21 through a serial port interface 46 that is coupled to the system bus, but may be collected by other interfaces, such as a parallel port, game port or a universal serial bus (USB). A monitor 47 or other type of display device is also connected to the system bus 23 via an interface, such as a video adapter 48. In addition to the monitor, personal computers typically include other peripheral output devices (not shown), such as speakers and printers.

[0043] The personal computer 20 may operate in a networked environment using logical connections to one or more remote computers, such as a remote computer 49. The remote computer 49 may be another personal computer, a server, a router, a network PC, a peer device or other common network node, and typically includes many or all of the elements described above relative to the personal computer 20, although only a memory storage device 50 has been illustrated in FIG. 1. The logical connections depicted in FIG. 1 include a local area network (LAN) 51 and a wide area network (WAN) 52. Such networking environments are commonplace in offices, enterprise-wide computer networks, intranets and the Internet.

[0044] When used in a LAN networking environment, the personal computer 20 is connected to the local network 51 through a network interface or adapter 53. When used in a WAN networking environment, the personal computer 20 typically includes a modem 54 or other means for establishing communications over the wide area network 52, such as the Internet. The modem 54, which may be internal or external, is connected to the system bus 23 via the serial port interface 46. In a networked environment, program modules depicted relative to the personal computer 20, or portions thereof, may be stored in the remote memory storage device. It will be appreciated that the network connections shown are exemplary and other means of establishing a communications link between the computers may be used.

[0045] Turning now to FIG. 2A, shown a multi-processor computing operating environment 100, which is suitable for practicing this invention and may be a particular embodiment of processing unit 21 (FIG. 1). Although FIG. 2A illustrates four (4) CPUs, any number of CPUs may be in the environment 100, including a single CPU. Each CPU 101, 102, 103 and 109 is connected to a memory controller 150 via bus 140 and includes cache memories L1 (110, 120, 130, 190) and L2 (115, 125, 135, 195). The memory controller 150 retrieves and stores data between RAM memory 155 and cache memories L1 (110, 120, 130, 190) and L2 (115, 125, 135, 195) as required by CPUs (101, 102, 103, 109).

[0046] A typical commercially available computer has extremely fast L1 cache memory (110, 120, 130, 190) which accommodates 8K bytes of code plus 8K bytes of data, and 512K bytes of fast L2 cache memory (115, 125, 135, 195). The L1 (110, 120, 130, 190) and L2 (115, 125, 135, 195) cache memories are divided into 32 byte cache lines. A cache line being the smallest unit transferred in and out of cache memories (110, 115, 120, 125, 130, 135, 190, 195). When the required data is not available to a CPU in its L1 cache (110, 120, 130, 190), a cache line fault occurs and the data must be loaded from lower speed L2 cache memory (115, 125, 135, 195), or relatively slow RAM 155. Thus, the application is stalled during the loading of this data, and until such time as the data is available to the CPU. Not only does decreasing the number of cache faults cause an application to run faster, but it also decreases the traffic load on bus 140, which can be a computer system bottleneck, especially in multi-CPU systems.

[0047] Turning now to FIG. 3, the steps for optimizing cache line groupings of elements within structures are presented. First, in step 310, the source code of the application to be optimized is compiled with a compiler to add instrumentation so that a stream of data (see FIG. 4) will be generated that logs each access to each data structure element when the application is executed. In one embodiment, the compiler inserts application programming interface (API) calls within the application to code that implements the logging functionality described herein. To profile the accessing of data elements within structures, the compiler inserts the call _DLP_Profiling just at the start of every basic block in the program. _{—DLP_Profiling is defined as}

[0048] int *_DLP_Profiling(void *CurrentProcAddr, int Size),

[0049] where CurrentProcAddr is the address of the current function (the function into which this API has been inserted). The value of CurrentProcAddr is used to find the name of the function in the program database (i.e., the PDB file generated by the linker). The Size parameter is the amount of space in bytes required to store all of the memory references made in this basic block. The compiled application is then linked with data source and logger code for logging the data as described herein to produce a second version of the executable application. This application is then executed and used in a typical manner such that the generated data stream characterizes the accessing of data elements during a typical execution session of the application.

[0050] FIG. 4 illustrates a record 405 generated for each data element accessed during execution of the application. As shown, record 405 is composed of a 32-bit memory address 410; and 32 bits of type information 420 containing a R/W bit 422 indicating whether the data element was read or written, a 13-bit offset variable 424 corresponding to the offset value of the data element within its structure; and an 18-bit definition variable 426 indicating the structure to which the data element belongs. In keeping with the scope and spirit of the invention, this data could have been written in a numerous different formats (e.g., individual data records rather than using bit fields as described herein for preserving space) and additional information could have been recorded. Moreover, the sizes of these bit fields will vary among compilers employed and computer systems in which the invention is practiced.

[0051] To be able to write these data elements at the rate at which they are produced, the optimal data logging method disclosed herein is used. Referring to FIG. 2B, shown is a high-level hardware block diagram representing the optimal data logging method. In one embodiment, sixteen 4 Kbyte memory pages X₁-X₁₆(201, 202, 216) within the system memory 22 of computer system 20 (FIG. 1) are filled with data records 405 (FIG. 4). Within each memory page(201, 202, 216), the data records 405 are sequentially stored at increasing memory addresses. Once full, memory pages X₁-X₁₆(201, 202, 216) are collected, and an array of pointers and a count of array entries are passed to a SCSI controller 210, which writes the memory pages (201, 202, 216) to the hard drive 220 using the WriteFileGather operation (described in FIG. 7A) on a computer system 20 running the Windows NT operating system. The software processing is described in detail herein with reference to FIGS. 5A-5C, 6A-6B and 7A-7B. Although this invention is described in the context of a personal computer 20 running the Windows NT operating system, this invention, could be practiced on any computer platform and operating system. In addition, the functionality of the WriteFileGather operation could be provided as an operating system service, by another application, or implemented as part of the data logger itself. Furthermore, multiple data sources, multiple data loggers, multiple hardware devices, and different disk logging techniques such as disk striping can be incorporated into the methods and systems disclosed herein in keeping with the scope and spirit of the invention.

[0052] Turning first to FIG. 5A, shown is a high-level view of the software architecture for optimally logging data produced by a data source 430 which is in a separate process than the data logger 434 which manages the writing of the data to a hard drive or other hardware device. As illustrated, the present invention can have multiple data sources 430 and/or multiple data loggers 434. This optimal data logging method uses two sets of buffers, the empty buffer pool 436 and the full buffer queue 432. These sets of buffers can be implemented as linked list structures as illustrated in FIGS. 5B-5C, arrays, or by using a multitude of other data structures commonly known in the computer arts. Similarly, the optimal data logging architecture presented in FIG. 5A can be readily expanded to incorporate multiple data sources 430, and multiple data loggers 434 which could efficiently store the data on multiple devices.

[0053] Next, in FIG. 6A, a flow diagram for the data source 430 (FIG. 5A) is presented. In step 462, a pointer to an empty buffer (442, 444, 446) (FIG. 5B) from the empty buffer pool 436 (FIGS. 5A, 5B) is retrieved. Next, in step 464, memory pages referenced within the retrieved buffer are sequentially filled with the data to be logged. After the buffer is full, a pointer to the newly filled buffer is then placed at the end of the full buffer queue 432 (FIGS. 5A, 5C) to maintain the ordering of the data. The data source 430, next in step 468, determines whether there is more data to be written. If so, this process is repeated.

[0054] FIG. 6B provides a flow diagram for the data logger 434 (FIG. 5A), which is operating in a separate thread from that of the data source 430. In step 472, the full buffer queue 432 (FIGS. 5A, 5C) is examined to see if there are any buffers to write. When there are buffers in the queue of full buffers, processing continues with step 474 in which buffers are retrieved from the front of the full buffer queue 432 (FIGS. 5A, 5C). Next, in step 475, these buffers are consolidated and packaged for writing to the hardware device. In the Windows NT environment, sixteen buffers (corresponding to 64 Kbytes or sixteen 4 Kbyte memory pages) are packaged into a larger data block by filling an array with pointers to memory pages retrieved from the full buffer queue 432. In one embodiment, better overall logging performance is achieved by limiting the data block to contain no more than one-half of the total number of buffers. After which, the writing of the packaged larger data block is performed in step 476. This writing of the data is performed using direct memory accesses (DMA). Such DMA memory accesses do not involve the computer CPU, and allow data transfer directly between memory and a peripheral device, such as a disk drive. In the Windows NT environment, the WriteFileGather command, described in FIG. 7A, logs the data to a hard disk after being passed the larger data block and the number of array entries which then. In other computing environments in which this invention is practiced, the logging process can use a command provided by the operating system of that environment or implement the functionality within itself. After the write operation is completed, step 478 is executed to determine if the logging process should end. If not, then the steps of illustrate by this flow diagram are repeated.

[0055] Returning to FIG. 3, in step 320, the voluminous data stream is processed to determine an optimized ordering of data elements within each structure as explained with reference to FIGS. 5-9. FIG. 9 provides a flow diagram of the processing of step 320. In step 610, the Hash Table 500 and array 510 represented in FIG. 8 are populated with the collected data for pairs of data elements within the same instance of a structure accessed within some number of operations of each other.

[0056] Turning to FIG. 8, Hash Table 500 provides the hashing table entry point with the individual elements of the hash table stored in Array 510 in a conventional linked list manner using Pointer (PTR) 550. A hashing function is used because of the sparsely populated nature of the pairings of the data elements. In one embodiment, Hash Table 500 is sized to be one-eighth the size of the computer's virtual address space; and the hashing function uses the concatenation of Address₁and Address₂as its seed. As would be evident to one skilled in the art, numerous other hash table sizes and hashing functions could be used, or some other data storage technique used.

[0057] Looking at array 510, each entry has seven data elements in addition to the Pointer 550. First, Address₁(Addr₁) and Address₂(Addr₂) store the addresses of the first and second element in the pair of elements plotted. Type Index₁and Type Index₂correspond with Type Index 420 (FIG. 4) from the data stream for the data element stored at the Address₁and Address₂, respectively. Counts are kept for all read and write operations between the data elements stored at Address₁and Address₂. For efficiency while retaining additional information, forward and backward counts are kept for read and write operations, and Address₁is defined as always being less than Address₂in Array 510. If, however, in sequencing through the data stream and a reference was made to a data element with a higher address, then the backward respective counter is used.

[0058] Pseudo code for populating the Hash Table 500 and Array 510 according to step 610 is provided in FIG. 10. For each element in the stream, the next n_neighborsubsequent elements are added to the Hash Table 500 and Array 510. In one embodiment, n_neighborhas been set to 256, corresponding to 8K (bytes L1 cache memory) divided by 32 (bytes per cache line). However, the value of neighbor can vary depending on the application being optimized. For example, a larger number might be used if a large number of data items are already allocated to the appropriate cache line; or possibly a smaller number would be better is a large amount of outside data (e.g., _imp_foo data or other compiler generated memory references) is being touched between accesses to data elements within the same structure definitions.

[0059] After all data elements in the stream have been added in accordance with the invention, the elements in Array 510 are processed to determine the optimized cache line allocation. Returning to FIG. 9, step 620 is performed to order the array 510 sorting (key 1) by the Address₁in ascending order; (key 2) by total (forward plus backward) writes in ascending order; and (key 3) by total (forward plus backward) reads in descending order. By sorting first by Address₁, all entries in the Array 510 for each structure are continuous as conventional compilers require structures to occupy a single block of memory. Finally, sorting by the second and third keys order the entries for any given Address₁such that the entries with the most writes propagate down, and those with the most reads propagate upwards.

[0060] Next, in step 630, the possible cache line combinations for each structure are generated to form a series of linear equations as described in the pseudo code of FIG. 11. For any given value of Address₁in each structure, a determination is made whether read or write operations predominate the accessing of the potential cache line allocation. A calculation is made to determine if the total number of write operations for the value of Address₁is less than one-third of the total number of read operations, then the cache line is composed of predominately written data elements, else by predominately read data elements. Then, a linear equation is comprised by selecting data elements from the bottom or top of those elements with the value of Address₁depending on whether write or read operations respectively predominate.

[0061] An illustration of the operation of the pseudo code of FIG. 11 is provided in FIG. 12, which shows a series of entries for two different values of Address₁for a single structure. First, for the entries with Address₁equal to A, because write operations predominate (i.e., total writes is not less than one-third the total reads), enough data elements-are selected to fill a cache line from the bottom of the sub-list 800. The selected data elements are shown within box 870. Therefore, a linear equation will include the data elements A, E, D, and F. Also, this equation will have a weighting of 27, the number of data accesses to these data elements. In other words, (ADEF)=0+(4+7)+(3+3)+(1+9)=27.

[0062] Next, referring to sub-list 830 for Address₁equal to B, as read operations predominate, the linear equation is formed from elements at the top of sub-list 830. This linear equation will include the data elements B, I, F, and G. This equation will have a weighting of 49, the number of data accesses to these data elements. In other words, (BFGI)=0+(20+0)+(18+2)+(6+3)=49.

[0063] Once all the linear equations for a structure are derived, these equations are solved in a conventional manner well known in the computer and mathematical arts as indicated in step 640 of FIG. 9. Once these equations are solved, an optimized grouping of data elements will have been determined for each structure. Continuing with the example presented in FIG. 12, the two derived potential cache line groupings were ADEF and BFGI, which both contain the element F. Therefore, only one of these groupings can be used (i.e., a data element can only be in one cache line). Because (BFGI)> (ADEF), (i.e. 49>27), the grouping BFGI is selected, and new potential grouping of ACDE will be selected from sub-list 800.

[0064] The solution of these linear equations provide the preferred optimized cache line groupings for the data structures of the application. The source code of the application is then edited to reflect this ordering as indicated in step 330 of FIG. 3. This editing can either be done manually or automatically, and the changes can be set-up as conditional definitions (e.g., using # ifdef statements) such that a switch can be set to direct the compiler to use the original or optimized structure definitions. The application is then compiled and linked per step 340, resulting in the optimized application which can be executed.

[0065] In view of the many possible embodiments to which the principles of our invention may be applied, it should be recognized that the embodiment described herein with respect to the drawing figures is only illustrative and should not be taken as limiting the scope of the invention. To the contrary, the invention as described herein contemplates all such embodiments as may come within the scope of the following claims and equivalents thereof.