Method for Efficient Non-Virtual Main Memory Management

The invention relates in general to operating system software for computer processing systems. More particularly, the present invention relates to methods for managing main memory for a non-virtual memory computer system.

Present day computer systems normally have an operating system, usually software interfacing with hardware, which functions to perform operations essential to the successful running of the computer system. The operating system for computer systems of the type incorporating the present invention have a storage management system which consists of complex software and hardware mechanisms. Computer systems of this type can perform multiprogramming and can include multiple tightly coupled parallel processors sharing a common main storage or memory. Because of this, the storage management system, for the successful operation of the computer system, must ensure that the common main storage is available for use without undue delay by the multiple programs and processors. One technique used by the storage management system to accomplish this is process or program swapping. When more programs or processes are in various states of execution than can be stored in main storage, the storage management system transfers one or more processes to a secondary storage device, called a swap device, such as disk storage or other type of mass storage, i.e. swapping out, and an executable process can be transferred from the swap device into main storage by the storage management system, i.e. swapping in.

Also, some operating systems have a storage management system which can do demand paging to facilitate multiprogramming. In such computer systems, the entire process or program does not have to reside in main storage to execute. The storage management system loads pages of a process on demand at the time the process references the required pages.

The present invention is incorporated in a computer system that does not have demand paging. The entire process to be executed must reside in main storage and remain therein until it becomes ineligible to execute. Therefore, while the storage management system of some computer systems can not perform demand paging to facilitate multiprogramming or the use of highly parallel multiprocessors sharing a common main storage, they can use process swapping. Some computer systems using swapping, such as Unix.TM..sup.1 System V, select swapout candidates according to their priority and residence time in main storage. The entire process image of the selected candidate is swapped out. In some instances the amount of free main storage space required for the process to be swapped in causes the swapping out of more than one process; however, using the same selection parameters. This results in greater swapping overhead and a less efficient and slower storage management system than the present invention because the present invention involves location and size when selecting swapout candidates.

In the present invention in addition to a list of available memory segments, separate lists of allocated and of free memory spaces are provided. The available and allocated circular lists are disjoint lists of segments which together account for all of memory. These lists are ordered by size. A third circular list orders all of the segments, both allocated and available, by their location in memory. This enhances efficiency because segment size and location can now also be used as parameters when selecting swapout candidates.

Additionally, past storage management systems, such as the storage management system of UNIX System V did not have a data structure which embraced capabilities of the present invention such as segment splitting. The present invention improves the efficiency and speed of the storage management system by providing the segment splitting capability. Without the segment splitting capability, prior storage management systems had to swap out enough contiguous segments from main storage which would make room for the total size of the segment to be swapped in.

The shared image data structure of the present invention enables main storage management to perform its function for microprocessing which was also lacking in prior storage management systems.

The present invention is an improvement over such past storage management systems in that it selects the least number of the most inefficient processes as candidates for swapout. Additionally, the present invention provides for the swapping of a subset of the total process image, i.e. partial swapping. Partial swapping allows the mixing of pieces of very large processes in main storage with smaller processes. This is accomplished without causing undue system overhead such as in the past when large processes were completely swapped out. These improvements provide a more efficient and faster main memory management system.

The principal objective of the present invention is to provide an improved main memory management system for computer systems and more particularly for non-virtual computer systems which support multiprogramming and can include multiple tightly coupled processors sharing a common main storage, and still more particularly to provide an improved main storage management system for such computer systems which is more efficient and faster.

These objects are achieved by incorporating process swapping into the main storage management system where the swapout candidates are selected according to the least number of the most inefficient processes and where a subset of a total process image can be selected for swapout.

New data structures are provided so that the main memory management system can perform improved swapping. These new data structures in the present invention take the form of lists and tables. One table, the segment table, is made up of segment entries which contain fields and flags which indicate logical boundaries, physical addresses, segment splits, and fields indicating if a segment is resident in memory, on a swap device, or both. One circular list, savail, orders available memory segments by size. One circular list, sactive, orders allocated memory segments by size. One circular list, sloc, orders both available and allocated memory segments by location. The savail, sactive, and sloc lists are threads through the segment table entries. One table, the shared image process table, contains fields and flags which indicate the status of process images.

In the preferred embodiment, i.e. non-virtual computer systems which support multiprogramming and can include multiple tightly coupled processors sharing a common main storage, each processor at a minimum has one set and preferably more, of data control registers which are used by the memory management system to map segment splits thereby enabling a user process to view a split segment as being contiguous, i.e. unsplit. By doing segment splitting and segment coalescing, available memory space can be used without reconfiguring memory, i.e. using main memory as it is currently allocated. When split segments are swapped out of main memory, they are re-united or coalesced on the swap device to allow the segment to be split into unique fragments to again fit the current memory configuration at the time of the next swapin.

Also, tables contain a flag used to facilitate partial swapping. This flag indicates if a segment of a process is in main memory or has been swapped out. By this arrangement, if a process to be swapped in can fit in a segment of a process which is a candidate to be swapped out, then it is only necessary to swap out the segment rather than the entire process which has been done in the past. Partial swapping can also accommodate a segment which needs additional memory space in order to grow beyond adjacent free space. Partial swapping to allow growth may be more efficient than the past method of memory to memory segment copy depending on the difference between memory-to-memory and memory-to-swap-device transfer rates of the hardware the memory management system is being used on.

When a process has some segments in main memory and some of its segments are on the swap device, the segments already in main memory are dual image segments. Therefore, when that process is being swapped in, only the segments which are not already in main memory are swapped in rather than the entire process image. Dual images of a process's segments are kept as long as possible so as to reduce the amount of memory allocation and physical I/O required to move an entire process image into main memory.

Although in the present invention main memory management is implemented in a computer system of the type described in the above referenced patent applications, it is not limited to this implementation. The method of the present invention can be implemented in other functions and utilities that require a main memory management scheme for a non-virtual or high speed, highly parallel, multiprocessor computer system.

These and other objectives of the present invention will become apparent with reference to the drawings, the detailed descriptions of the preferred embodiment and the appended claims.

Coalescing of memory is the process by which fragmented data segments are merged into one contiguous space when they are swapped out.

Data control register sets are hardware registers used to map fragmented data segments so that the user sees one contiguous segment.

Dual images are processes which have both reserved main memory and a swap image at the same time.

Fork is a standard UNIX system call invoked by a parent process in order to create a new child process.

Microprocesses (mprocs) are a type of lightweight process that have a very low context switch overhead, are discardable upon exit, are reusable prior to disposal, provide a means for dividing up work to be done into very small segments, and return only the result of the work with no other context information.

Partial swapping allows a portion of a large process to be swapped out for another process that is to be swapped in or in order to grow beyond adjacent free memory.

Process image is the representation of the process resources within the operating system, such as context information, executable code, and data for a process.

Process table contains control and status information needed by the kernel about each process in the system.

Sactive is a dummy segment structure which heads a circular doubly linked list of allocated segment entries ordered by size.

Savail is a dummy segment structure which heads a circular doubly linked list of available segment entries ordered by size.

Segment table is made up of segment entries which describe attributes of sections of memory which they govern, such as whether they contain text or data, whether they are shared or private, whether they are split or unsplit, and where the data of each segment is located in memory and/or on a swap device.

Segments are contiguous parts of a process image which are required to be in main memory during the process's execution.

Shared image is a single process image possibly shared by multiple microprocesses executing multiple threads of a multithreaded program.

Shared image process table is made up of entries which contain information about a process image and about segments belonging to that image. If the image is not shared, it is a single process's image. If microprocesses are sharing the image, this entry contains information related to those microprocesses.

Shared memory is a type of data segment which can be split, written to, and shared by more than one process, allowing direct communication between the sharing processes.

Sloc is a dummy segment structure which heads a circular doubly linked list of allocated and available segment entries ordered by location.

Split is a fragment of a segment which resides in main memory and is described by a segment table entry.

Swap device is an area in an alternate memory used to temporarily hold a process image not currently running in order to free main memory for a process about to run. The swap device is commonly a block device in a configurable section of a disk.

Swapping is the act of allocating segments and moving the data in the segments between main memory and a swap device.

FIG. 1 is a schematic block diagram of a single multiprocessor cluster system incorporating the present invention.

FIG. 2 comprised of 2(a) and 2(b) are diagrams illustrating the sofware architecture for the multiprocessor system of FIG. 1.

FIG. 3 is a block diagram illustrating the swapin and swapout functions.

FIG. 4 is a block diagram of relevant fields in the segment table.

FIG. 5 is a diagram illustrating memory layout lists.

FIG. 6 is a block diagram of relevant fields in the shared image process table.

FIG. 7 is a diagram illustrating the relationship between process table entries, shared image process table entries, shared memory segments, and split segments.

FIG. 8 is a diagram illustrating the dual image concept.

FIG. 9 is a diagram illustrating the splitting of a data segment.

FIG. 10 is a diagram illustrating memory layout as pertains to choosing a swapout candidate.

FIG. 11 is a diagram illustrating the coalescing of a segment.

FIG. 12 is a diagram illustrating the standard non-virtual UNIX method of creating a child process via the swapout procedure when enough main memory is not available.

Appendix A contains the Memory Management Pseudo-code for the preferred embodiment of the present invention.

With reference to the drawings and particularly to FIG. 1, the present invention by way of example is illustrated as being incorporated in a single multiprocessor cluster system having an operating system kernel 1100 resident in main memory 14. FIG. 1 is substantially the same as FIG. 5 of the referenced commonly assigned co-pending patent application entitled Integrated Software Architecture for a Highly Parallel Multiprocessor System, Ser. No. 07/537,466, filed on Jun. 11, 1990.

FIG. 2 illustrates the operating system kernel 1100 which includes a memory scheduler or main memory management system 1120. FIG. 2 is substantially the same as FIG. 8 of the aforementioned referenced patent application. The main memory management system 1120 functions to allocate storage space in main memory 14 for processes to be run and de-allocate storage space in main memory 14 from those processes which have run and are in a wait state. Main memory 14 has a finite amount of storage space and normally there are more processes in the system than can be stored in main memory 14. Memory management 1120 is responsible for finding or making main memory 14 space available to the process which is to be currently run by the computer system. A process image is a representation of the process resources within the operating system, such as code, data, stack, context information, and registers. This process image is separated into a multiple number of segments of type text, data, shared memory, library text, library data, or user. If space for this process image is unavailable in main memory 14, the memory manager 1120, FIG. 3, can swapout an entire process image, multiple process images, or a subset of the segments within a process image and place it on a mass storage device such as an alternate memory, a disk storage system, or other suitable storage device which is termed a swap device such as swap device 70. The process image to be run is residing on swap device 70 and it is swapped into main memory 14. That process will then be executed by a processor 10 either running to completion or going into a wait state due to some condition such as waiting for an input/output, I/O, device. While in this wait state the process becomes a candidate for swapout.

The memory manager 1120 in the preferred embodiment of the invention uses the segment table 90, of which a partial entry is shown in FIG. 4, to keep track of segments belonging to processes in the system. Fields which are discussed herein and which are most relevant to the functioning of the memory manager are included in the diagram. A segment table 90 entry is created and maintained by the memory manager 1120 for each segment and each segment split belonging to a process which is in the system.

The memory manager 1120 in the preferred embodiment of the invention uses the three circular doubly linked lists in FIG. 5 to keep track of space in main memory 14. The sloc 20 list orders all main memory 14 segments, whether free 50 or allocated 60, by location. Sloc 20 uses pointers kept in each segment table 90 entry (s.sub.-- floc is the forward link, s.sub.-- bloc is the backward link) to link all segments. The savail 30 list orders main memory 14 segments which are free 50, or available for allocation. The sactive 40 list orders main memory 14 segments which are currently active 60, or unallocatable. Both the sactive 40 and the savail 30 lists are ordered by segment size. Both sactive 40 and savail 30 uses pointers kept in each segment table 90 entry (s.sub.-- fsz is the forward link, s.sub.-- bsz is the backward link) to link segments. Entries in savail 30 and sactive 40 are mutually exclusive, as are use of the segment table 90 pointers s.sub.-- fsz and s.sub.-- bsz. The sum of the entries in savail 30 and sactive 40 equal the entries in sloc 20. The memory manager 1120 updates these lists to reflect the current memory 14 configuration each time memory 14 is reconfigured due to processes growing or entering or exiting the system. The memory manager 1120 uses these lists for allocation and freeing of main memory 14.

The memory manager 1120 in the preferred embodiment of the invention uses the shared image process table 190, of which a partial entry is shown in FIG. 6, to keep track of which processes are using each segment, whether segment splitting is possible, and information related to swapout candidates. Fields which are discussed herein and which are most relevant to the functioning of the memory manager 1120 are included in the diagram. A shared image process table 190 entry exists for every process image in the system. Although the memory manager 1120 maintains a number of fields in the shared image process table 190, the remaining fields are maintained and utilized by other parts of the operating system kernel 1100.

The relationship between the process table 180 entries, shared image process table 190 entries, and segment table 90 entries are shown in FIG. 7. In this example, process table 180 entries 1 and 2 are actually microprocesses sharing text 1 and the split data 1. Microprocesses are thoroughly described in the referenced commonly assigned co-pending patent application entitled Dual Level Scheduling of Processes to Multiple Parallel Regions of a Multithreaded Program on a Tightly Coupled Multiprocessor Computer System, Ser. No. 07/571,958. Process table 180 entry 3 is a process which is sharing text 1 with microprocesses 1 and 2. Their shared image process table 190 entries are linked via the shared text link (si.sub.-- shtx.sub.-- link) to indicate this. Process table 180 entry 3 is sharing data 2 with process table entry 4. Their shared image process table 190 entries are linked via the shared data link (si.sub.-- shmm.sub.-- link) to indicate this. Process table 180 entry 5 represents a process that is not sharing any segments in its image.

The following functional description of the memory manager 1120 relates directly to the memory management pseudo-code provided in Appendix A. The functional description references the uses of the tables and fields within the tables previously mentioned.

As in standard UNIX, as part of the operating system kernel 1100 initialization, the sched process (often referred to as swapper), is initiated to allocate main memory 14 to all processes which are ready to execute, but whose process images reside on the swap device 70 and are not yet loaded in main memory 14. When there are no process images on the swap device 70 which are ready to execute, the memory manager 1120 is put in a sleep state which suspends its execution until the operating system kernel 1100 resumes its execution both periodically and when memory allocation is needed. Sched loops through the process table 180 entries FIG. 7, searching for the most eligible process which is in the "ready to run" state, but is not loaded in main memory 14. Commonly, most eligible is defined as the oldest process which is waiting only for main memory 14, some implementations also use priority as a criteria for most eligible; alternately, most eligible can take on any definition as long as no stipulations (except for the allocation of main memory 14) prevent immediate execution. Regardless of the definition of eligibility, fields in the process table 180, such as "base level user priority" and "recent processor usage", are used to define which process is the most eligible.

In order for a process to execute, its entire process image must reside in main memory 14. Each segment belonging to a process image is described in a segment table 90 entry. Each segment table 90 entry contains an s.sub.-- size: field which indicates the amount of main memory 14 required for that segment. If the required space in main memory 14 is obtained for all segments of this process image, these segments are copied from the swap device 70 to main memory 14, that is, swapped in. This mechanism is accomplished by a routine called swapin, which will be explained hereafter. If enough space in main memory 14 for all segments is not obtainable, the memory space which was obtained is freed and allocation is attempted again. This allows split segments to be coalesced, then resplit in a new way, possibly fitting in the free space 50, FIG. 5 (splitting and coalescing are described in more detail hereafter). If memory allocation is still not possible, sched will put itself in the sleep state. The clock routine will wake sched at the next major clock tick at which time sched will then retry the swapin. This delay provides a chance for main memory 14 to reconfigure before allocation is attempted again. Reconfiguration of a shared main memory 14 may occur during this delay due to the movement of processes in memory 14 which are running on other processors. If swapin still fails to allocate space in memory 14, space in main memory 14 is forcibly reconfigured before swapin is attempted again.

Forcible reconfiguration is achieved by swapping out the non-running process images residing sequentially in memory 14 until enough main memory 14 is freed for all of the eligible process image segments; that is, as a last resort, the first swappable segment on the sloc 20 list is swapped out to force a memory reconfiguration. If the swapin succeeds, the next most eligible candidate to be swapped in is selected from the process table 180 for main memory allocation. If swapin was still unsuccessful, sched is yet again put in the sleep state until the next major clock tick to allow memory 14 to reconfigure. This loop continues until all swapped runnable processes are swapped in. It should be noted that sched looping through the process table 180 occurs as in standard UNIX, but the method of acquiring space in main memory 14 is unique to the present invention.

Swapin is the procedure called from sched which finds allocatable or free main memory space 50 for each segment belonging to a process image. Although the swapin concept exists in standard UNIX, the implementation thereof described here is unique to the present invention. Swapin determines main memory 14 layout by using the sloc 20, savail 30, and sactive 40 lists. For each segment in a process image, swapin begins by checking if the segment already exists in main memory 14; this is indicated by the segment table field s.sub.-- flags containing a value of SG.sub.-- LOAD, FIG. 4.

FIG. 8 shows one such situation referred to as dual images, where segments of a process image exist (are stored) in main memory 14 while also existing (being stored) on the swap device 70. Dual images result when swapin fails to acquire enough main memory 14 for all of a process image's segments. The process retains the space in memory 14 it did obtain for the process image's segments in the hope of finishing the memory allocation on the next attempt for it to be swapped in. Until this process does acquire enough main memory 14 for all of its segments, it is defined to be a process with a dual image. It is only when enough space in main memory 14 is obtained and all of a process image's segments are copied into main memory 14 from the swap device 70, that the swap image is freed. Dual images are prime candidates for swapout, since their space in memory 14 need only be freed, not copied, because the swap images have not yet been cleared from the swap device 70.

Partial swapping would also create a situation where swapin would encounter some segments of a process in main memory 14 and other segments not in memory 14. Partial swapping is the process whereby only a subset of the total number of segments of a process image have been swapped out. Partial swapping can be done to allow another process's segments to be swapped in. Partial swapping can also be done if a single segment of a process is swapped out in order to grow. Swapping out to allow growth is a standard real memory (i.e. non-paged) method of growing a segment in memory 14 when it cannot be grown in place. That is, if space in memory 14 following this segment is already allocated (i.e. allocated memory 60, FIG. 5) and is therefore not available for this segment's growth, then the segment is swapped out, its size adjusted to include the desired growth, and it is swapped back in at its new size in a free memory space 50 that is large enough to contain it. Some memory management systems accomplish this growth by moving the segment to a larger area of free memory space 50, but in the present invention, it is more efficient to swap the segment rather than do a memory 14 to memory 14 copy, due to the particular hardware platform having a block transfer rate which is faster than the word to word memory transfer rate.

Regardless of whether the situation was created by the existence of dual images or from a partial swap, if the segment data to be swapped in already exists in main memory 14, no further allocation processing is required for it. If the segment to be swapped in does not have allocated space in main memory 14, a single memory block of sufficient size to hold the entire segment is sought by scanning the savail 30 list. If a single free space 50 in memory 14 of at least the proper size does not exist, segment splitting is considered.

Segment splitting can be accomplished if the segment to be swapped in is splittable and if the processor 10 where this process will execute has unused data control register sets 80, FIG. 9. Data control registers are thoroughly described in the referenced commonly assigned pending patent application entitled Global Registers for a Multiprocessor System, Ser. No. 07/536,198. Commonly, only data segments are defined to be splittable, but the implementation of the preferred embodiment accepts any alternate definition. The data control register sets 80 are entities of the hardware system, belonging to each processor 10. Each data control register set 80 consists of three control registers (the DSEGxS, the DSEGxE, and the DSEGxD registers, where x is between 0 and the maximum number of data segments) used to map logical to physical addresses. Each data control register set 80 records the lower bound (logical starting address) 100, the upper bound (logical ending address) 110 and the displacement 115 which must be added to the logical address to obtain the physical address 120. The lower bound 100, upper bound 110, and actual physical starting address 120, are recorded in the segment table 90 entries in fields named s.sub.-- lb, s.sub.-- ub, and s.sub.-- base, as listed in FIG. 4. The data control register sets 80 map the splits to enable the user to interpret the fragments of the segment as a contiguous segment. The number of data control register sets 80 per processor 10 determines the maximum number of splits per process image. That is, for each split there must be a set of registers to map a user's logical address to its physical location. A field in the shared image process table 190 FIG. 6 entry (si.sub.-- data.sub.-- avail), keeps track of how many data control register sets 80 remain unused and can therefore be used to map splits. If it is determined that the current segment can be split, the largest piece of free memory 50 is reserved and the bounds thereof and displacement are entered in one of the data control register sets 80 to describe split. The SG.sub.-- SPLIT value is entered in the segment table 90 entry s.sub.-- flags field to indicate that this is a split segment. The upper bound 100, lower bound 110, and physical address 120 description are saved in the segment table 90 entry for this split (fields s.sub.-- ub, s.sub.-- lb, s.sub.-- base). The segment table 90 entry for this split is also updated to include links to the previous and next split of this segment (fields s.sub.-- prvsplt and s.sub.-- nxtsplt). The shared image process table 190 entry is adjusted to hold the current count of remaining splits available (field si.sub.-- data.sub.-- avail). As long as unused data control register sets 80 for this processor 10 remain and the segment to be swapped in is not completely in memory 14, this splitting process continues.

In the case that there is no free memory space 50 for the current nonsplittable segment, intelligent choices are made about how to create free space 50. The six choices for swapout candidates follow in order of priority. First, segments belonging to shared process images that already have some of their segments moved to the swap device 70 are swapped out. This includes dual image segments and partially swapped images. Second, segments belonging to sleeping processes 150 which with adjacent free space 50 are of sufficient size for the current segment are swapped out. Third, segments belonging to adjacent sleeping processes 150 who together free enough space for the current segment are swapped out. Fourth, segments belonging to both runnable 160 and sleeping 150 processes who together with adjacent free space 50 are of sufficient size for the current segment are swapped out. Fifth, segments belonging to adjacent runnable 160 and sleeping 150 candidates who together free enough space for the current region are swapped out. Finally, as a last resort, shared memory segments are swapped out. FIG. 10 illustrates a main memory layout, including free space 50, swapout candidates 150 and 160, and non-candidates 170 for the swapout procedure previously described. That is, all of the segments under the heavy arrow in the diagram could be swapped out if necessary to free enough main memory 14 space for an incoming segment.

When memory is allocated for a segment by the swapin procedure, the value SG.sub.-- JUST.sub.-- IN is set in the s.sub.-- flags field of the segment table entry 90 FIG. 4 to keep track of which segments need physical copying. That is, segments already containing data do not need to be copied in from the swap device 70; their s.sub.-- flags field indicates this by containing the value SG.sub.-- LOAD. This situation occurs when a process is a dual image or is partially swapped out. For example, if only the data segment of a process was swapped out, then only that data segment is copied back to memory 14 from the swap device 70. The count of allocated (in-memory) segments in the shared image process table 190 FIG. 6 entry field si.sub.-- segs.sub.-- in is also incremented at this time.

When the count of allocated segments is equal to the total number of segments for this process image, space in memory 14 has been allocated for the entire image so segments on the swap device can be copied into memory 14. The segments are sequentially processed. This processing includes locking the segment table entry to make it inaccessible by other processes, physically copying the data of size s.sub.-- size contained in the segment from the swap device 70 at the address saved in the segment table 90 entry field s.sub.-- swapaddr 70 to main memory 14 at the address saved in the segment table 90 entry field s.sub.-- base, and setting the value SG.sub.-- LOAD in the s.sub.-- flags field of the segment table 90 entry (indicating memory 14 now contains the segment data).

If any of the segments just swapped in are shared memory segments (those being shared between more than one process) each shared image process table 190 entry is updated. This includes updating reference counts and marking that process image as loaded if all segments for that process image are now in memory 14.

At the successful completion of memory allocation for the current process, it is able to execute. Sched then continues searching the process table 180 for another eligible process to swapin, continuing the loop described earlier.

Out.sub.-- seg is the procedure called during swapin which creates free space 50 in main memory 14 by moving data in a segment from memory 14 to a swap device 70. The particular segments to be swapped out are determined as previously described.

Initially, the segment table entry of the segment whose contents are to be swapped out is locked to make it inaccessible by other processes.

If the entries in the segment table 90 FIG. 4 indicate that this segment has been split into more than one piece (s.sub.-- flags contains the value SG.sub.-- SPLIT), then those splits are re-united or coalesced. Coalesing, which implies that contiguous space is allocated on the swap device 70 for all splits and the contents of the splits are merged into this area when written from memory 14, is shown in FIG. 11.

The segment, split or unsplit, of size s.sub.-- size is physically copied from main memory 14 at the address saved in the segment table 90 entry field s.sub.-- base to the swap device 70 at the address saved in the segment table 90 entry field s.sub.-- swapaddr 70. Then, depending on exact circumstances, flags are updated to reflect the swapout status. The shared image process table 190 entry reference count (si.sub.-- segs.sub.-- in), that is the count of the number of segments in memory 14, is decremented for each image sharing this segment. The shared image process table 190 entry field si.sub.-- data.sub.-- avail is adjusted to reflect the coalesing of the splits. The values SG.sub.-- SPLIT, SG.sub.-- JUST.sub.-- IN, and SG.sub.-- LOAD are cleared from the segment table entry field s.sub.-- flags to indicate that it is no longer split, memory is no longer reserved for it, and data for it is no longer contained in main memory 14. The data control register sets 80 describing this segment are also cleared.

If swapout is being called as the result of a fork (indicating a parent process image is being copied to the swap device 70 without destroying the parent process image to create a new child process image), main memory 14 for this segment is not freed. In all other cases, space in main memory 14 for the swapped out segments of a process is freed. In the case of a fork, the child process image on the swap device 70 will be swapped in later when the child process is eligible to execute. This standard real memory (non-paged) creation method, is illustrated in FIG. 12.

Although the description of the preferred embodiment has been presented, it is contemplated that various changes could be made without deviating from the spirit of the present invention. Accordingly, it is intended that the scope of the present invention be dictated by the appended claims rather than by the description of the preferred embodiment. ##SPC1##