Method and system for enhanced concurrency in a computing environment

Abstract

A system and method for performing an operation within limits upon a shared value stored in an actual value location. The method provides a plurality of memory locations, two of the memory locations being first and second reservation memory locations, and two of the memory locations being limit memory locations containing limit values. An operand is obtained and the operation is performed upon an affected reservation register using the operand. A resulting value is compared to the limit values in the limit memory locations and if the resulting value is not within the values in the limit memory locations, an inverse operation is performed to restore the affected reservation register and a failure is reported. Else the operation is performed to update the actual value location, the operation is performed to update an unaffected reservation register, and a success is reported.

Description

BACKGROUND OF THE INVENTION

DESCRIPTION OF THE BACKGROUND ART

[0002] The present invention relates generally to multitasking computer systems, and more particularly, to multiprocessor systems and a method for increased concurrency in such systems.

[0003] Multiprocessor computing systems have been used for years to perform multiple tasks. A multiprocessing computing system may be defined as a computer assembled from a plurality of independently or semi-independently operating intelligent, i.e., processor-based, stations. The stations are generally interconnected for communication by a communication bus. Each processor may perform a plurality of tasks with apparent concurrency. Distributing the system intelligence among a plurality of stations may improve the processing throughput of the computing system, allow the system to modularly and uniformly expand to meet increasing computing requirements, and permit greater efficiency and productivity through concurrency.

[0004] While multiprocessor computing systems are known, they are not scalable. The software of such systems is often highly dependent upon the configuration and characteristics of the system hardware. This means that the scalability of the multiprocessor is limited. Further, the dependency of the software on the hardware often causes a waste in processor resources.

[0005] Synchronizing or coordinating the tasks using a central processor to enhance concurrency of a plurality of tasks has been problematic in most multiprocessor systems. The prior patents of Sullivan, U.S. Pat. Nos. 4,484,262, 4,707,781 and 5,438,680, the disclosures of which are incorporated herein by reference, addressed some issues relating to enhancing concurrency.

[0006] The Sullivan patents describe that concurrency can be enhanced by communicating between processors or tasks via commonly accessible memory to reduce conflicts within the multiprocessor system. However, the disclosed techniques are primarily dependent on hardware, such as the processors.

[0007] Uninterruptible or atomic operations are operations that cannot be interrupted by another processor or by another thread on the same processor. Intel processors, for example, perform addition as an atomic machine instruction with extremely short locking implemented by a cache coherency mechanism. A LOCK prefix implements the atomic operation for a set of operations that do read-modify-write operations on a single memory address such as the LOCK XADD instruction (Exchange Add). The lock time of atomic instructions is generally very short.

[0008] Current technology suffers from the inability to add or subtract from a shared memory location without locking the location during the operation if it is necessary to keep the value within limits or maintain a valid transaction log. Using atomic operations, it is possible to add or subtract without locking which causes blocking, provided that there are no limits and a valid transaction log is not necessary.

[0009] Therefore, a need exists for a system and method for enhancing concurrent computation with shared resources. Further, a need exists for a software-based system that is capable of enhancing the concurrency of any multiprocessor computing system.

SUMMARY OF THE INVENTION

[0010] In one aspect, the invention is directed to a method for operating a computer system. The computer includes at least one processor. The method includes establishing a plurality of memory units each having a corresponding memory location. A plurality of tasks running on the processor are executed, and the plurality of tasks are operable to share data. A plurality of lists for each memory location are defined, and at least one of said lists is locked if the data is invalid. An entry is inserted into said locked list corresponding to one of said tasks. The locked list is unlocked, and a determination is made if data is inputted in the memory location between the determining step and the unlocking step.

[0011] In another aspect, the invention is directed to a method for operating a computer system. The system includes at least one processor. The method includes establishing a plurality of memory units each having a corresponding memory location. A plurality of tasks are run on said processor, and the plurality of tasks are operable to share data. An entry corresponding to one of the tasks is inserted into one of the lists if the list is unlocked. The method also includes determining if another of said lists is unlocked if said one list is locked.

[0012] In another aspect, the invention is directed to a method for synchronizing processes in a computer system. The computer system includes at least one processor. The method includes establishing a plurality of memory units each having a corresponding memory location. A plurality of tasks running on the processor are executed, and the plurality of tasks are operable to share data located in the memory units. A plurality of lists for each memory location are defined. A list is locked if the data is not valid. An entry is inserted into the locked list corresponding to one of the tasks, and the list is unlocked. The entered task is suspended until valid data is found in the memory unit. The method also includes the steps of reading valid data and determining if other data appears in said memory location before said locking step and after said unlocking step. The other data is read if it appears in the memory unit.

[0013] Implementations of the invention include one or more of the following. The locking step further comprises activating selected other ones of said plurality of tasks that are entered on said locked lists. The plurality of lists may form a linked list. The plurality of lists may be between four and eight. The method may also include the step of transferring the operation of said locked list when said locked list is locked by another one of said plurality of tasks.

[0014] In another aspect, the invention is directed to a computer system having enhanced concurrency that includes a plurality of processors. A plurality of tasks are run on the plurality of processors. The computer system may also include a plurality of memory units each having a corresponding memory location, and a plurality of lists corresponding to each of said memory locations. One of said tasks may be responsible for activating selected ones of said plurality of tasks contained on the same list as said one task.

[0015] In yet another aspect, the invention provides a system and method for adding and subtracting within limits without blocking. A thread comprises an operation including either addition of an addend or subtraction, which is performed as addition of a negative of the addend. The operation affects a shared actual value stored in an actual value register. Upper and lower limit registers store permissible upper and lower limits within which the result of the operation must fall. Addition and subtraction reservation registers, which are also shared resources, store a reserved value after an addition to the actual value or after a subtraction from the actual value respectively.

[0016] The method includes getting the value of the addend. If the operation is addition the value of the addend is positive, otherwise it is negative. A LOCK XADD operation using the addend is performed on an affected reservation register, i.e., the addition reservation register if the operation is addition or the subtraction reservation register if the operation is subtraction. The resulting value in the affected reservation register is then compared to the value of the limit registers. If the operation cannot succeed, the addend is added back to the affected reservation register in a LOCK XADD operation and a failure is reported.

[0017] If the operation can succeed, in a LOCK XADD operation the addend is added to the value of the actual value register. Finally, in a LOCK XADD operation, the addend is added to the value of the unaffected reservation register which was not affected by the first atomic operation and a success is reported.

[0018] Aspects of the invention include a computer-implemented method for adding and subtracting within limits without blocking. Another aspect of the invention provides a system having a central processing unit (CPU) and a memory which are configured to effect the method described above. Another aspect includes a computer program recorded on a computer readable medium for causing a computer to effect the method as described above.

[0019] The foregoing and other objects and advantages of the disclosed system and method will become apparent to those of ordinary skill in the art after having read the following detailed description of the preferred embodiments that are illustrated in the various drawing figures.

BRIEF DESCRIPTION OF THE DRAWINGS

[0020] FIGS. 1A-1D are flowcharts illustrating a preferred method of reading data from a mailbox.

[0021] FIGS. 2A-2D are flowcharts illustrating a preferred method of inputting data in a mailbox.

[0022] FIG. 3 is a flowchart of a “one last look” method.

[0023] FIGS. 4-24 illustrate additional aspects of the invention.

[0024] FIG. 25 illustrates a portion of a conventional computer system, including a CPU and a memory, in which the present invention may be embodied.

[0025] FIG. 26 illustrates a system block diagram in accordance with a preferred embodiment of the invention.

[0026] FIG. 27 illustrates the overall process for implementing a method for adding and subtracting within limits without locking.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0027] The present invention is directed to a system and method for enhancing concurrency in a multiprocessor or multitasking computer system. A multitasking computer system may be defined as a computer system that permits multiple threads of execution in a single execution space. A multiprocessing computer system may be defined as a computer system that allows multiple processing units to share data. Preferably, various tasks in the computer system communicate using commonly accessible “mailboxes.” A mailbox is an area of memory that contains a number of states, and may be identified by a virtual address. When one task produces an item of valid data that may be desired by multiple other tasks, the task holding the valid data inputs it into a mailbox. Generally, other tasks read the valid data from the mailbox.

[0028] In accordance with a preferred method, the task that inputs the valid data into the mailbox is caused to notify other tasks addressing the mailbox that the data is contained therein. This means that there is no need for central coordination of mailbox accesses. Further, this minimizes busy waits for valid data. This permits tasks and processors to allocate their resources more productively.

[0029] As stated above, the mailbox is a memory location. The mailbox includes a pointer to the address corresponding to valid data. Alternatively, a number, such as a string of zeros, is employed if the data is invalid. For simplicity, the preferred method will be described with reference to inputting or reading data to or from the mailbox. However, the data to be accessed can be located in the same memory system as the mailbox. Further, the term “task” refers to a set of instructions running on a process. Thus, the processor executes the task. Additionally, the terms “invalid” and “valid” refer to arbitrary logical status. In the preferred method, a task reads valid data and does not read invalid data. Other states may be used such as “true” and “false”.

[0030] A plurality of lists of tasks wait for valid data associated with the mailbox. By coordinating several lists in accordance with the preferred method, conflicts in accessing data and delays in obtaining data can be minimized. This is because only a small number of instructions, for example, under 30, are required to read or place data to or from the mailbox. These functions occupy a small percentage of a task's activity time, such as 5% or less. Accordingly, it has been found that a smaller number of lists can eliminate most conflicts. However, a list for each active task may be required to eliminate all conflicts. Any number of lists may be used. For example, the number of lists may be between four and eight.

[0031] In the preferred method, the lists are a linked list. A linked list can employ locations in the same memory system as the mailboxes. The linked list includes locations having suitable space for at least two pointers. The locations in the middle of the list may contain a forward pointer to the next location including the list, and a backward pointer to the preceding location including the list. Additionally, the top location on the list may contain a pointer forward to the next location and a pointer backward to an illegal address, such as zero. The bottom location on the list can contain a pointer backward to the preceding location and a forward pointer to an illegal address, such as zero. When an item, i.e., data, is added to the list, the forward pointer corresponding to the previous bottom location is changed to correspond to the location of the added item. Also, the added item becomes the new bottom location. When the top item is taken from the list, the next item in the list replaces the top item. Additionally, the backward pointer of the added item is changed to point to an illegal address, such as zero. When both pointers of an item in a mailbox are to an invalid address, the list is considered empty.

[0032] Reading Valid Data When Available

[0033] FIGS. 1A-1D are flow charts showing a preferred method for reading data from a mailbox. Initially, data in a mailbox is read (step 101). In the preferred method, the mailbox may contain a pointer to the memory location containing the desired valid data or an illegal address corresponding to invalid data, such as zero. The mailbox is then addressed and interrogated to determine if the data is valid (step 102). At this stage, a mailbox may contain or is about to contain data associated with a program of a task seeking such data. If valid data is in the mailbox, it is read from the memory address indicated in the mailbox (step 104). The task may also read the data during step 102.

[0034] One aspect described herein is that the task performing steps 101 and 102 contacts other tasks waiting for valid data from that mailbox. This means that conflicts are avoided in the multiprocessor system because busy waits for valid data are minimized.

[0035] FIGS. 1A-1D illustrate lists L0 and L1 that represent tasks waiting to receive valid data from a mailbox, and FIGS. 2A-2D shows lists L2 an dL3 that identify tasks waiting to input valid data in the mailbox.

[0036] As shown in FIG. 1A, a task list is locked when it is manipulated by a task. This prevents access by other tasks to that list. After list L0 is locked (step 106), the task determines if list L0 was previously locked by another task (step 108). If list L0 was previously locked by another task prior to the current task executing steps 106 or 108, the prior task operates list L0. The current task then proceeds to check list L1 (step 122), as described below.

[0037] If list L0 was previously locked (step 108), the current task determines if list L0 is empty (step 110). If the list L0 is not empty, the current task then removes the first waiting task (“waiter”) from the list (step 116). A waiter may be defined as a task that waits for valid data to be in a mailbox. In this configuration, a waiter may be in a suspended state, as described below. The current task then unlocks list L0 such that other tasks may access the list (step 118). Additionally, the current task causes the task associated with that waiter to execute a wake code (step 120). A wake code may be defined as a set of computer instructions that cause a task to activate from a suspended mode. Then, the current task again locks list L0 (step 106). This process continues until each task on list L0 has been notified, and the list L0 is empty. Alternatively, this process may continue until the list L0 has been found locked. If this list is empty, list L0 is unlocked (step 112), and the task proceeds to check list L1 (step 122), as described below.

[0038] If list L0 was previously locked (step 108), the current task locks list L1 (step 122), and then checks if that list was previously locked (step 124). If list L1 was not previously locked at step 124, the current task determines if list L1 is empty (step 128). If list L1 is not empty, the current task removes the waiters from list L1 (step 132) and causes the waiters to execute their wake codes (step 134), as described above. This process repeats until list L1 is empty or the list L1 has been found locked. When list L1 is empty, it is unlocked (step 130). The task then continues to execute its remaining instructions as defined by the system (step 126). If list L1 was previously locked by another task prior to steps 122, 124 or 126, that task operates the locked list. The current task then executes step 126.

[0039] Valid Data Not Available

[0040] If no valid data is found at the mailbox (step 102), list L0 is locked (step 140). The current task then determines if list L0 was previously locked (step 142). If the list was not previously locked, the current task is entered on list L0 (step 144). List L0 is then unlocked (step 146), and the current task checks for valid data (step 148). The current tasks then determines the validity of the data (step 150). This above process may be referred to as “one last look,” as described below. In the preferred method, this configuration keeps two or more tasks, which are synchronized by passing data through the mailbox, from missing each other.

[0041] If the data is still not valid at step 150, the task is suspended (step 152). This releases computer system resources for other tasks, or the task can complete other system instructions. This means that the task does not busy wait or periodically check for valid data in the memory location. The current task will be activated by another task which has found valid data in the memory location, or by a task which previously inputted the valid data (step 154). When activated by another task, the current task proceeds to read the valid data from the mailbox (step 104). If the data is valid (step 150) and list L0 was previously locked (steps 156 and 158), the current task is suspended until activated (steps 152 and 154). When activated by another task, the current task proceeds to read the valid data from the mailbox (step 104).

[0042] If the data was valid (step 150) and the current task determines that list L0 was not locked (step 156 and 158), the current task reads the data. This may occur as the current task determines the validity of the data. Then, the current task executes step 202.

[0043] If the current task is on the list L0 at step 202, the current task is removed from the list (step 204). The next waiting task is also removed from the list L0 (step 206), and list L0 is again unlocked (step 208). The task then causes the other task removed from the list to execute its wake code (step 210).

[0044] Next, in accordance with the preferred method, the current task determines if the list L0 was previously locked (steps 212 and 214). If so, the current task then executes other system instructions (step 220). If the list was not previously locked, the current task then determines if the list is empty (step 216).

[0045] If list L0 is not empty (step 216), the current task loops back to remove and notify the next task at the top of the list (steps 206-216). If on any such loop, the list is found locked at steps 214 or 216, the current task continues to perform other system instructions (step 220). When the list is empty, it is unlocked (step 218) and the current task continues to perform other system instructions (step 220).

[0046] The current task may not be on the list L0 (step 202). This means that another task has removed the current task from the list and is operating to activate it. However, the current task also operates list L0. Thus, the current task must activate the other tasks on list L0. The current task may then suspend itself until activated by another task.

[0047] Steps 228 through 238 operate similar to steps 206 through 218. However, when list L0 is found locked (step 234) or empty (step 236), the current task is suspended (step 240). When activated by another task (step 242), the current task proceeds to read the valid data from the mailbox (step 104).

[0048] Referring again to FIG. 1B, if list L0 was locked (step 142), the current task proceeds as follows. List L1 is locked and then the current task determines whether that list was previously locked (steps 180 and 182). If list L1 was previously locked, the current task returns to lock list L0 (step 140), and continues to search for an unlocked list. This example provides two possible lists. However, more lists may be used. If the number of lists equals the number of tasks, no task will repeat this loop.

[0049] If list L1 was not previously locked (step 182), the current task becomes a waiter on list L1 (step 184), the list L1 is unlocked (step 186), and a “one last look” check is made for valid data (steps 188 and 190). If the data is not valid, the current task is suspended (step 152).

[0050] If the data is valid (step 190), then the current task reads the data as it confirms the validity of the data. List L1 is then locked (step 192), and a check is made whether it was previously locked (step 194). If list L1 was previously locked (step 194), list L1 is unlocked (step 196), and the current task is suspended (step 152).

[0051] If the data was valid and it is determined that list L1 was not locked, at steps 190, 194, and 196, respectively, then the current task proceeds (step 252).

[0052] FIG. 1D shows that if the current task is on list L1 (step 252) that task is removed from the list (step 254), and also removes the next waiting task from the list L1 (step 256). The task then unlocks list L1 again (step 258). The task then causes the other task removed from the list to execute its wake code (step 260).

[0053] Next, the current task determines whether the list L1 was previously locked (steps 262 and 264). If it was not previously locked, the current task determines whether the list is empty (step 266).

[0054] If list L1 is not empty (step 266), the current task loops back to activates task at the top of the list (steps 256 through 266). If on any such loop, the list is found locked at steps 264 or 266, the current task continues step (270). When the list is empty, it is unlocked (step 261), and the current task continues to execute other system instructions (step 270).

[0055] If the current task is not found on the list L1 (step 252), this means that some other task has taken the current task off the list and is attempting to activate it. The current task also operates list L1. Thus, this task must activate the other tasks on list L1 before suspending itself. The current task can then be activated by another task. The method of steps 278 through 288 is similar to steps 256 through 268 described above. However, when list L1 is found locked (step 284) or empty (step 286), the current task is suspended (step 290). When activated by another task (step 292), the current task proceeds to read the valid data from the mailbox at step 104.

[0056] This preferred method of finding an available list is simple. Further, a task should not be significantly delayed in finding an unlocked list. Thus, enhanced concurrency can be achieved over prior systems.

[0057] Inputting Valid Data

[0058] FIGS. 2A-2D are flow charts showing the input of data into a mailbox in accordance with the preferred method. Initially, data is read in a mailbox (step 301). The mailbox is then addressed and interrogated to determine whether the data is valid (step 302). If valid data is not already in that mailbox, the mailbox is marked as not containing valid data. Next, the valid data is placed in the mailbox (step 304).

[0059] In accordance with the preferred method, lists L2 and L3 identify tasks waiting to input valid data in the mailbox. This includes a current task waiting to input data in the mailbox. The current task determines whether list L2 was previously locked (step 304). If so, that list remains locked (step 306). If list L2 was not previously locked the current task determines whether list L2 is empty (step 310). If the list L2 is not empty, then the current task removes the first waiter from that list (step 316). Also, the current task unlocks list L2 so that other tasks may access it (step 318), and causes the task associated with that waiter to execute its wake code (step 320). Then, the current task locks list L2 again (step 306). The above process continues from step 306 until all tasks on list L2 seeking the valid data from the mailbox have been activated. Alternatively, this process continues until the list L2 is empty, or the list L2 has been found locked. When list L2 is found empty, it is unlocked (step 312), and the current task proceeds to list L3 (step 322).

[0060] If list L2 was previously locked (step 301) or list L2 was found empty (step 310) and then unlocked (step 312), the current task then locks list L3 (step 322). Then, the current task checks if list L3 was previously locked (step 324). If list L3 was not previously locked, the current task determines whether list L3 is empty (step 328). If list L3 is not empty, the current task proceeds to remove the waiters from list L3 (steps 332 and 334) and causes them to execute their wake codes (step 306). This process repeats by looping back to step 322. The process may continue until list L3 is empty or has been found locked. When list L3 is empty, it is unlocked and the current task continues to execute other system instructions (steps 330 and 338). If list L3 is found to have been previously locked at steps 322 or 324, this indicates activity by another task. The current task then executes other system activities (step 338). The current task also defers to the other task to activate the other tasks listed on list L3.

[0061] Valid Data Found in Mailbox

[0062] If the current task determines that there is valid data at the memory location (step 302), the list L2 is locked (step 304). The current task also checks whether the list was previously locked (step 342). If list L2 was not previously locked, the current task is inserted onto list L2 (step 344). List L2 is then unlocked (step 346) and a “one last look” check is made for valid data (steps 348 and 350), as described below. If the data is still invalid, the current task is suspended (step 352). The current task is then activated by another task that has found or inputted valid data in the memory location (step 354).

[0063] If the data in the mailbox is valid and list L2 was previously locked (steps 356 and 358), the current task is suspended until activated (steps 352 and 354). When activated by another task, the current task proceeds to input valid data into the mailbox (step 304).

[0064] If the data was not valid (step 350) and list L2 was not previously locked (steps 356 and 368), the current task continues to execute other system commands (step 402).

[0065] If the current task is on list L2 (step 402) the current task removes itself from that list (step 404). Also, the current task removes the next waiting task from the list L2 (step 406) and unlocks list L2 again (step 408). The current task then causes the other task to be removed from the list to execute its wake code (step 410).

[0066] In accordance with the preferred method, the current task determines whether the list L2 was previously locked (steps 412 and 414). If so, the current task proceeds to input valid data in the mailbox (step 304). Otherwise, the current task determines whether the list is empty (step 416).

[0067] If list L2 is not empty (step 416), the current task loops back to remove and activate the next task at the top of the list (steps 406-416). If on any such loop, the list is locked at steps 414 or 416, the current task inputs valid data in the mailbox (step 304). When the list is empty, it is unlocked (step 418). The current task then inputs valid data in the mailbox (step 304).

[0068] If the task is not on the list L2 (step 402), the current task may be removed from the list and is being activated by another task. The current task also operates list L2. Thus, the current task must activate the other tasks which are on list L2 before suspending itself. The current task may then be activated by another task. The method of steps 428 through 438 is similar to steps 406-418. However, when list L2 is found previously locked (step 434) or empty (step 436), the current task is suspended (step 440). When activated by another task (step 442), the current task proceeds to input valid data in the mailbox (step 304).

[0069] If list L2 was previously locked at step 342, the current task locks list L3 (step 380) and then determines whether that list was previously locked (step 382). If list L3 was previously locked, the current task loops back to attempt to access list L2 (step 340), and continues to find an unlocked list. This example provides two possible lists. However, more than two lists may be employed.

[0070] If list L3 was not previously locked (step 382), the current task is inserted on list L3 as a waiter (step 384). The list L3 is also unlocked (step 386), and a “one last look” check is made for valid data (steps 388 and 390), as described below. If the data is valid, the current task is suspended (step 352).

[0071] If the data is not valid (step 390), list L3 is locked, (step 392) and a check is made whether list L3 was previously locked (step 394). If list L3 was previously locked, list L3 is unlocked (step 396), and the current task is suspended (step 352). If the data was not valid and the list L3 was not previously locked (steps 390, 394, and 396), the current task proceeds to step 452.

[0072] If the current task is on list L3 (step 452), it is removed from that list (step 454). Also, the next waiting task from the list L3 is removed (step 456), and the current task unlocks list L3 again (step 458). The current task then causes the other task removed from the list to execute its wake code (step 460).

[0073] The current task then determines whether the list L3 was previously locked (steps 462 and 464). If it was not previously locked, the current task determines whether the list is empty (step 466).

[0074] If list L3 is not empty, the current task loops back to remove and activate the next task at the top of the list (steps 456-466). If on any such loop, the list is locked (steps 464 and 466), the current task inputs valid data in the mailbox (steps 304). When the list is empty, it is unlocked (step 468), and the current task determines if valid data is inputted in the mailbox (step 304).

[0075] If the current task does not find itself on list L3 (step 452), this may be because some other task has taken the task off the list and is trying to activate it. The current task, however, operates list L3. Thus, the current task must activate other task on list L3 before suspending itself. The current task then may be activated by another task. The method of steps 478 through 488 is similar to steps 456 through 468. However, when list L3 is found previously locked (step 484) or empty (step 486), the current task is suspended (step 490). When activated by another task (step 492), the current task proceeds to input valid data in the mailbox (step 304).

[0076] One Last Look

[0077] The term “one last look” refers to the situation where a task that has placed itself on a list for valid data from a mailbox receives that data. The preferred method allows tasks addressing the same mailbox to be coordinated. This means that a current task inputting valid data in a mailbox will activate another task that was entered onto a wait list, while the data was placed in the mailbox by the current task. Accordingly, a current task may not input valid data in the mailbox, while another task is inserted on a list to receive that data. The preferred method avoids the potentially infinite period of time that a task may sleep until another task activates it. Thus, delays in system processing time may be minimized.

[0078] FIG. 3 shows that a task first looks at the data in the selected mailbox and determines whether it is valid data (steps 548 and 550). If the data is valid, it reads the data and continues with other system activities (steps 551 and 553).

[0079] At this stage, if the data is not valid, the task is inserted onto a wait list (step 555). Then, the “one last look” determines if the data is valid (steps 557 and 559). This occurs because, while the task that may have located an unlocked list, locked it, placed itself on the list and unlocked the list again, another task may have put valid data in the mailbox. Without this preferred “one last look” method, activating the inserted task may be delayed until another task seeks valid data from the mailbox.

[0080] If the data is not valid at that time, the task will suspend until activated by another task, such as the next task inputting valid data in the mailbox (steps 552 and 554). This method corresponds generally to steps 352 and 354.

[0081] If the data is valid at that time (step 559), the task can take itself off the list (step 561). Alternatively, the task may read the data and continue its other system activities. This means that the task may assume that the data was initially found in the mailbox (steps 563 and 553). A number of embodiments of the present invention have been described. Nevertheless, it will be understood that various modifications may be made without departing from the spirit and scope of the invention. For example, the number of lists may be increased to accommodate additional tasks. Accordingly, it is to be understood that the invention is not to be limited by the specific illustrated embodiments, but only by the scope of the appended claims.

[0082] Method For Synchronization of Multiple Asynchronous Threads Without Spin Locks, Busy Waits or Polling

[0083] Current technology does not possess the ability to synchronize multiple asynchronous threads without employing busy waits, or polling. This causes a waste of processor resources and limits the scalability of multi processors, because the likelihood of busy waits and the time consuming polling increases with the number of active threads. In addition, it is inefficient with current technology to multi thread programs in which a large number of elements effect the state and a large number of elements take the state as input (such as simulations).

[0084] An unlimited number of threads may interact without ever busy waiting or polling in an embodiment. It is possible to write programs which can execute efficiently on uni processors and arbitrarily large multi processors without modification.

[0085] A thread of execution may be defined as a sequence of instructions which are parsed out of memory and executed by the processing hardware and whose state is maintained on a single stack. A mailbox may be defined as an area of memory which contains at least two states (usually called the valid state and the invalid state). A multilist may be defined as a structure composed of two or more lists which logically act as one list.

[0086] Associated with this mailbox is at least one structure which can contain references to threads of execution which have been suspended pending on of the permitted states of the mailbox. One implementation of this structure is a linked list. Threads will contain instructions which reference this mailbox. One implementation of these instructions is as methods of a mailbox object. This instruction, when executed, tests the state of the mailbox, and if the state is one which permits the completion of the instruction (called the valid state), completes the instruction. If the state of the mailbox is invalid, the thread is placed on a list associated with the invalid state and the thread is suspended (this method is appropriate for both user level control and system level control. In this example only user level control will be explained. In user level implementations, the suspension is at user level and does not involve a system call). There is another kind of instruction which changes the state of the mailbox (Instructions may both wait on a state and change the state.) from invalid to valid. When this instruction is executed by any thread, all threads on the newly valid wait list cease to be suspended and are made runnable. The preferred methods may include many implementations of system level synchronization and event mechanisms.

[0087] A difficulty with any mechanism which suspends and resumes threads of execution is that a list must exist somewhere of the threads to be resumed and that more than one thread cannot be altering the list at any one time. This means that the list must be locked when it order is being altered. If another thread requires access to the list, it must wait its turn (busy wait). It does not help to suspend the thread because that just requires another identical “wait list”. It also does not help to pass the problem to a single thread because that also creates an identical list for the attention of the single thread (It also limits scalability by creating a limit on the number of processors which can perform a particular function). One embodiment includes several parts and solves the general case of access contention. Applications include not only synchronization, but also the general case of concurrent programming as exemplified by simulations, and by the non blocking access to databases by multiple read and write operations.

[0088] By using multiple lists (or queues etc.) to record the waiting threads and allowing any thread which finds its first choice of lists locked to use another one (and another if that one is locked), the chances of busy waiting can be made arbitrarily small for any thread trying to get itself on the list. The chance of busy waiting can even fall to zero if the number of lists is equal to the number of threads needing access. This may not be necessary because the probability of busy waiting declines faster than exponentially and approaches zero (becomes an insignificant barrier to scalability) for very small numbers of lists.

[0089] Accordingly, a solution to the input problem may be to use multiple lists.

[0090] If thread one finds the mailbox data invalid, and goes to put itself on a list, but thread two, after thread one has checked the data, makes the data valid and resumes all the threads on the list before thread one is even on the wait list, thread one will wait forever. This is a synchronization failure. The solution to this synchronization problem is that thread one must check back (“one last look”) on the mailbox data after it has placed itself on the list and released the list. If it finds that the data is valid, it must remove itself from the list and continue. Now it is impossible for the synchronization to fail.

[0091] To reduce or eliminate the chance that a busy wait will occur because the reader and writer of synchronizing data contend for the same list, since they must share access for the synchronization to have meaning, access contention may occur. The solution to this access contention is as follows:

[0092] A: A thread which is looking for its valid state, checks the mailbox data and if it is valid, returns that data and continues.

[0093] B: If that thread finds the data invalid, it finds an unlocked list (as above in part one) and locks it. It then places itself on the list, unlocks the list, and takes one last look at the mailbox data. If the data is still invalid the thread suspends itself.

[0094] C: If the data has become valid, the thread tries to lock the list on which it has placed itself. If the attempted lock fails, the thread suspends itself, it will be reawakened by whomever has the list locked.

[0095] D: If the lock succeeds, the thread removes itself from the list and causes all other members of the list to resume. The thread then continues.

[0096] 1: A thread which is creating a valid state for the mailbox, writes data indicating that state to the mailbox data area and then attempts to lock the first list of the multilist. If the lock succeeds, the threads waiting on the list are caused to resume. The thread which wrote the valid data then repeats (1 or 2) the procedure on the next list of the multilist.

[0097] 2: If the lock attempt fails, the list is skipped, and the one locking the list will cause all of the threads on the list to resume (see D). The thread which wrote the valid data then repeats (1 or 2) the procedure on the next list of the multilist.

[0098] 3: When the last list of the multilist is completed, the thread which wrote the data will continue.

[0099] This mechanism may guarantee that the synchronization mechanism will always be successful and that it will never busy wait. This introduces a fixed amount of overhead to the communication and coordination of threads which does not increase with the size of the program or the number of processors. This generally defines scalability.

[0100] It may be useful to permit synchronization on any Boolean. A Boolean “not” can be implemented by having a thread have as its valid state of a particular mailbox the complement of another thread's valid state (or zero and not zero etc.). A Boolean “or” can be implemented by having two or more threads write valid data to the same mailbox. A Boolean “and” can be implemented by having a single thread read two or more mailboxes in succession. With “not”, “and” and “or”, all Booleans can be encoded. This provides for the very useful control structure, “whenever (general Boolean condition), do (statement)” without polling.

[0101] One problem is making simulations and other general programs concurrent. A method, analogous to the synchronization mechanism, makes writing parallelizable simulation programs (and other inherently concurrent programs) easier. In this method:

[0102] A: The data is replaced by a list is associated with consecutive integers and which can be added to on one end only (an ordered list).

[0103] B: Multiple wait lists, ordered by the integer position being waited on, replace the multilist.

[0104] C: A request for data passes an integer and returns the data associated with that integer if the integer is lower than or equal to the end of the list. If it is higher than the number (any compare operator which divides the list at a point will do), the thread is placed on a wait list and the mechanism described above is used to resume it, except that the test of validity is whether the integer last written is equal to greater than the number requested.

[0105] Another problem making simulations and other general programs concurrent is that there is not always an obvious way to determine whether all segments of the program which might influence the inputs to a particular segment have executed. A method, analogous to the synchronization mechanism, makes writing parallelizable simulation programs (and other inherently concurrent programs) easier. In this method:

[0106] A: A map is made of the program, which models a space, in which regions may be defined as exclusively permitting the interaction of components in a particular circumstance. For instance, in a driving simulation, automobiles would only effect one another if they were in identical or adjacent regions of the street map. The objects which move in the model may be attached to an arbitrary number of regions, called interaction regions, by enrolling in lists such as described in continuation 2. Each region is required to report the transfers of object to its neighbors (a neighbor is a region which may accept an object from the region to which it is a neighbor). When the number of reports equals the number of regions for a given integer value, that integer value is current (able to execute knowing all inputs are present). A preferred implementation is to count with a fetch-decrement instruction and pass objects from region to region by using a list mail box.

Lock Free Record Management

[0107] There are two reasons for lock records. One is to keep it from being altered in inconsistent ways and the other is to protect it in some way. The first reason can be eliminated. Currently, locking dominates the cost of computing. Existing technology can do nothing to reduce this cost.

[0108] A mechanism for ordering operations to provide the same result as if there were locks without the need to block processes or threads is described. If the concurrency is high enough to overcome the inherent latency of access, a high percentage of optimal computing efficiency can be achieved. Accordingly, the number of necessary locks can be reduced.

[0109] A thread of execution may be defined as a sequence of instructions which are parsed out of memory and executed by the processing hardware and whose states is maintained on a single stack. A mailbox may be defined as area of memory which contains at least two states (usually called the valid state and the invalid state). A multilist can be defined as a structure composed of two or more lists which logically act as one list. Associated with this mailbox is at least one structure which can contain references to threads of execution which have been suspended pending one of the permitted states of the mailbox. One implementation of this structure is a linked list. Threads may contain instructions which reference this mailbox. One implementation of these instructions is a methods of a mailbox object. An instruction, when executed, tests the state of the mailbox, and if the state is one which permits the completion of the instruction (called the valid state), completes the instruction. If the state of the mailbox is invalid, the thread is placed on a list associated with the invalid state and the thread is suspended (This method is appropriate for both user level control and system level control. In this example only user level control will be explained. In user level implementations, the suspension is at user level and does not involve a system call). There is another kind of instruction, which changes the state of the mailbox (Instructions may both wait on a state and change the state.) from invalid to valid. When this instruction is executed by any thread, all threads on the newly valid wait list cease to be suspended and are made runnable. Many implementations of system level synchronization and event mechanisms can be included. A difficulty with any mechanism, which suspends and resumes threads of execution is that a list must exist somewhere of the threads to be resumed and that more than one thread cannot be altering the list at any one time. This means that the list must be locked when its order is being altered. If another thread requires access to the list, it must wait its turn (busy wait). It does not help to suspend the thread because that just requires another identical “wait list”. It also does not help to pass the problem to a single thread because that also creates an identical list for the attention of the single thread (It also limits scalability by creating a limit on the number of processors which can perform a particular function). Applications include not only synchronization, but also the general case of concurrent programming as exemplified by simulations, and by the non blocking access to databases by multiple read and write operations. Generally, the general case of access contention can be solved.

[0110] A read-modify-write operation that may be added (such as LOCK XADD on the Pentium) and can be used to distribute locally unique consecutive numbers. Operations that must be performed in order are assigned numbers such that the operation that must be done first has a lower number. Each operator that can access the database can place commands in a list sorted by number. So long as each operation executes on individual records in the correct order, the result must be the same. That this must be true is obvious from the fact that all locks do is guarantee some instruction will occur before some others.

[0111] A necessary instruction may not be delivered to the operation until an operation that needed to occur after it had already been executed in the above read-modify-write operation. A synchronizing mechanism should exist. Instead of locks, a block delivery system can be used. Instructions that must be executed in order are delivered in blocks. The synchronizing mechanism can be mailboxes.

Method for Adding and Subtracting within Limits without Blocking

[0112] Current technology suffers from the inability to add or subtract from a shared memory location without locking the location during the operation if it is necessary to keep the value within limits or maintain a valid transaction log. Using the uninterruptible operations common to most general purpose CPUs, it is possible to add or subtract without blocking, provided that there are no limits and a valid transaction log is not necessary.

[0113] It has been found that an unlimited number of threads to add and subtract from a single shared memory location without blocking, overflow, or underflow. Further, the transaction log can always be valid so that error recovery is possible.

[0114] A thread of execution may be defined as a sequence of instructions which are parsed out of memory and executed by the processing hardware and whose state is maintained on a single stack. A shared value can be characterized as an area of memory that is accessible from at least two threads of execution. An uninterruptible instruction may be an instruction that reads an area of memory and does not permit another thread of execution to read or write that same area of memory until the instruction has completed. A limit can be defined as a value that the contents of a memory location must not pass (such as zero for inventory items or the end of allocated memory space for a pointer).

[0115] A single memory location can be included that may be accessed only by an interruptible add or subtract operation such as LOCK XADD (exchange and add) on the Pentium or a nondestructive operation such as a read. Such a device may safely permit many threads to add or subtract, but it is impossible to determine whether incomplete transactions have occurred or to execute instructions, such as extended precision, that require synchronization.

[0116] One more memory location (for a total of two) may be added that can be accessed by uninterruptible instructions only, whatever limits exist (limits may exist in one direction only) and a transaction log. Now an addition or subtraction is accomplished by adding (a negative number for subtraction) to one of the locations (each thread must use the two locations in the same order), if the operation does not place the total on the wrong side of the limits, then the operation is recorded in the transaction log and then added to the second memory location. If the total violates the limit (is smaller than the lower limit or larger than the upper limit) the number is subtracted from the first number. The failure may be recorded in a transaction log. Generally, this is an implementation detail. This device can guarantee a correct result, and that the total includes only valid operations, but it leaves a significant chance that a valid transaction will not occur.

[0117] A preferred embodiment of the invention is practiced in the context of a conventional personal computer such as an IBM compatible personal computer, an Apple Macintosh computer or a UNIX-based workstation. FIG. 25 shows a representative hardware configuration of a computer 2500 including a central processing unit (CPU) 2510, such as a microprocessor, and a number of other units interconnected via a system bus 2520. The computer 2500 may also include Read Only Memory (ROM) 2540, Random Access Memory (RAM) 2550, Non-Volatile Memory 2560, input devices 2570 (such as a keyboard, mouse, microphone, and touch screen) and output devices 2580 (such as a display screen, printer, and speaker) coupled to the system bus 2520. A Network Connection 2590 may be provided for connecting computer 2500 to a communication network (not shown) such as an intranet or the Internet. The coomputer 2500 typically operates under control of an operating system such as the Microsoft Windows NT or Windows/98 OS, IBM OS/2, MAC OS, or UNIX operating system. Those skilled in the art will appreciate that the disclosed system and method can also be implemented on platforms and operating systems other than those mentioned.

[0118] With reference to FIG. 26, a system for adding within limits without blocking is shown generally designated 2600. The system includes an actual value register 2610, a subtraction reservation register 2620, an addition reservation register 2630, a lower limit register 2640, and a upper limit register 2650. RAM 2550 may include the registers 2610, 2620, 2630, 2640, and 2650.

[0119] A thread of execution including an atomic addition or subtraction operation (subtraction as addition of a negative value of an addend) is operable to operate upon the value stored in actual value register 2610. By way of example, actual value register 2610 is shown to have a value of 7 which may include the value of current inventory.

[0120] Subtraction and addition reservation registers 2620 and 2630 store a reserved value including the value of the actual value register 2610 after the operation. Thus for addition, the value of addition reservation register 2630 is the actual value increased by the value of the addend and for subtraction the value of subtraction reservation register 2620 is the value of the actual value decreased by the value of the addend.

[0121] Limit registers 2640 and 2650 store the value of the limits within which the operation is constrained. In the exemplary case of inventory control, the lower limit register 2640 is shown to have a value of 3 and the upper limit register 2650 is shown to have a value of 10.

[0122] With reference to FIG. 27, the method of the invention generally designated 2700 comprises a step 2710 in which the value of the addend, −3 in the exemplary case indicating a request of 3 items of inventory, is obtained. If the operation is addition, the value of the addend is positive, otherwise it is negative.

[0123] In a step 2720 a LOCK XADD operation is performed upon the value stored in an affected reservation register using the addend. If the operation is addition, the affected reservation register is the addition reservation register 2630, otherwise it is the subtraction reservation register 2620. Thus in the exemplary case, the addend −3 is added to the subtraction reservation register 2620, resulting in a value of 4 in the subtraction reservation register 2620.

[0124] In a step 2730, the value of the affected reservation register is compared to the value of the limit registers 2540 and 2550. In a step 2740, it is determined if the operation can succeed within the limits. In the exemplary case, the value of the subtraction reservation register 2620 is 4, which is greater than or equal to the lower limit of 3 stored in lower limit register 2540 and less than or equal to the upper limit of 10 stored in the upper limit register 2650. Therefore the operation can succeed within the limits.

[0125] If the operation cannot succeed, in a step 2750, the value of the affected reservation register is restored by performing a LOCK XADD using the negative of the value. Thus in the example, 3 is added back to the value of 4 stored in the subtraction reservation register 2620 to restore the initial value of 7 stored in the subtraction reservation register 2620. Finally, a failure is reported in a step 2760 and the process ends.

[0126] If the operation can succeed, in a LOCK XADD operation 2770, the addend is added to the value stored in the actual value register 2610 and written to the actual value register 2610. In a step 2780, in a LOCK XADD operation, the addend is added to the reservation register unaffected by the first LOCK XADD operation of step 2720. In the exemplary case, the unaffected reservation register is the addition reservation register 2660 and thus −3 is added to 7 and loaded into the addition reservation register 2660. Finally, in a step 2790, a success is reported and the process ends.

[0127] A wait list with an uninterruptible counter may be added. Now, a thread of execution checks that the value could be added to the second memory location within the limit, performs an uninterruptible add on the first location as above, and if it succeeds, then it proceeds as described above. If it fails the test against the limit, the second value I checked. If the operation would succeed on the second value (the operation to test the second value cannot alter it), the thread calculates the chances of success of each member of the wait list, in order, ignoring all members marked as causing overflow and marking all members that cause overflow. If the value of the second memory location plus the members of the list make it impossible to add the current value in without overflow, the value is subtracted back out of the first memory location and an overflow failure is reported. If the value can be succeed, the thread adds one to the counter and places itself in the location that the counter indicates (such as value times address length plus offset) with space for a read modify write operation. The thread then suspends for a time long enough to allow the last thread to successfully complete an operation to have a chance to mark the reserved location to indicate that the thread should continue. This time should be short enough to be otherwise benign. When the thread is reactivated it marks the space reserved for read-modify-write operation that it is done, if it has not been otherwise marked, then it subtracts the value it added from the first memory location. The last thread then marks members of the list that best exhaust the remaining space to the limit (application depended) and goes on. This device guarantees that an optimal result will be achieved without significant overhead and without much blocking, but it does not guarantee a unique last thread. It is possible that two threads will try to finish allocating the remaining value to the limit, and it is possible that some threads waiting to be serviced will be missed.

[0128] “One last look” and synchronization configurations may also be added. For implementation, two locations are added to hold semaphores. A semaphore may be defined as a memory location only written to by uninterruptible instructions. One case may be considered. When a value greater than one is added to the first total and causes the limit to be exceeded, but the second total is or has some room left before hitting the limit. There must be a semaphore in a known state (say reset). The thread that set the first value past the limit changes the semaphore (say sets it). It then subtracts its value from the first memory location. If the result is within limits, the thread then checks the counter. If the counter is zero, it resets the semaphore. If the counter is non-zero, it adds the value of each thread on the list to the second total. Because the semaphore blocks other writing done after the semaphore, and the first counter was known to be in bounds after the semaphore was set, there can be no case of overflow. If the first counter is out of bounds, values on the list are subtracted from the first counter until the counter is in bounds, then proceed as before. Threads arriving at the device check the semaphore. If it is reset, they proceed as in part three. If it is set, the value of the second memory location is checked. If it does not allow the transaction, a failure is returned. If the second memory location has a value that allows the transaction to complete, then the value is added to the first location and the value to be added is placed on the list and the thread is paused. When it awakens, if it marked as added, it returns success. If it is marked failed, it returns failure. If it is unmarked (meaning that is miss-connected with the thread that set the semaphore), then it subtracts its value from the first memory location and begins again. Because contention occurs only when a large value exceeds the limit because other values were processed between the check of the second value and the addition to the first value and smaller values are waiting to be processed, contention almost never occurs. Locks are otherwise unnecessary. An incorrect result may never occur.

Lock Free Lists

[0129] With current methods, it is necessary to lock a list to add or delete a member. It has been found that elements can be placed and removed without locks.

[0130] The lists may have many of the useful properties of queues without locks. These devices may also replace linked lists without locks.

[0131] The device may include a memory location and a contiguous block of memory. The memory location and the block are initialized a known value (like zero). Any thread of execution that would add to the list creates an array in memory and then adds one to the memory location with an uninterruptible instruction (such as XADD on the Pentium). From the value returned by the uninterruptible instruction, an address in the contiguous block of memory is computed. A pointer may be placed there pointing to the array just created. A thread can remove an element from the list by reading down the list and exchanging the first nonzero element with zero. This will be a pointer to an element of the list. If the thread trying to remove an element reads past the current write pointer, the list is empty.

[0132] The contiguous block of memory may be divided into multiple parts. Each part may have a memory location for a down counter and for a write location pointer. Each also may have a lowest unread pointer. In the rare case that a thread is suspended after it has gotten a write location and before it has written, locations after it on the list will be removed before it. The lowest unread pointer that allows the search to start as far down the list as possible without missing anything. When the down counter, which is decremented on each removal, is zero, the list may be reused. This permits the list to act like a queue, but without the locking of the head and tail pointers. (it scales).

[0133] A possible sequence may be implemented. The initializing process sets all of the memory to some known set of values that cannot be valid data. The down counters are set to the number of elements in a list section. The memory locations are set to the head of their respective sections of the list. All of the memory locations that point to the head of a contiguous portion of memory are placed in a list. The top of the list is pointed to in a master location.

[0134] The writing process increments the first section pointer by the block size. The XADD instruction returns the old value of the pointer, which is the top of the list. If writes into this location the pointer that it wishes to place in the queue. The process continues until the pointer equals the endpoint of the section. Then the writing process finds the first section that is empty (down-counter equals zero), sets the down-counter to the section length and places the address of the section pointer in the master pointer. The next writer uses the next block referred to by the counter.

[0135] The removing process reads down the section from the top until it finds a non zero element. It exchanges that element with zero. If a non-zero value is returned, that element is removed and the responsibility of the remover. Otherwise, it continues to scan down until it succeeds in getting by uninterruptible exchange a non-zero value. The removing process then decrements the down-counter associated with the section in question and if the down counter equals zero, marks the section empty and available for re-use.

[0136] Several embodiments are specifically illustrated and/or described herein. However, it will be appreciated that modifications and variations are covered by the above teachings and within the scope of the appended claims without departing from the spirit and intended scope thereof.

Claims

1. A method of operating a computer system, said computer system including at least one processor, comprising. establishing a plurality of memory units each having a corresponding memory location; executing a plurality of tasks running on said at least one processor, said plurality of tasks being operable to share data; defining a plurality of lists for each memory location; determining the validity of said data in said memory unit; locking at least one of said plurality of lists if said data is invalid; inserting an entry corresponding to one of said plurality of tasks onto said locked list; unlocking said locked list; and determining if data is inputted in said memory location between said determining step and said unlocking step.

2. A method for operating a computer system, said computer system comprising at least one process, comprising: establishing a plurality of memory units each having a corresponding memory location; running a plurality of tasks on said processor, said plurality of tasks being operable to share data; defining a plurality of lists for each memory location; inserting an entry corresponding to one of said plurality of tasks onto one of said plurality of lists if said one list is unlocked; and determining if another of said lists is unlocked if said one list is locked.

3. A method for synchronizing processes in a computer system, said computer system including at least one processor, comprising: establishing a plurality of memory units each having a corresponding memory location executing a plurality of tasks running on said processor, said plurality of tasks being operable to share data located in said memory units; defining a plurality of lists for each memory location; locking at least one of said plurality of lists if said data is not valid; inserting an entry corresponding to one of said plurality of tasks onto said locked list; unlocking said locked list; suspending said entered task until valid data is found in said memory unit; reading said valid data; determining if other data is inputted in said memory unit before said locking step and after said unlocking step; and reading said other data if it appears in said memory unit.

4. The method of claim 3, wherein the locking step further comprises activating selected other ones of said plurality of tasks that are entered on said locked list.

5. The method of claim 3, wherein said plurality of lists forms a linked list.

6. The method of claim 3, wherein said plurality of lists is between four and eight.

7. The method of claim 3, further comprising transferring the operation of said locked list when said locked list is locked by another one of said plurality of tasks.

8. A computer system having enhanced concurrency, comprising: a plurality of processors; a plurality of tasks running on said plurality of processors; a plurality of memory units each having a corresponding memory location; a plurality of lists corresponding to each of said memory location; wherein one of said plurality of tasks is responsible for activating selected ones of said plurality of tasks contained on the same list as said one task.

9. The system of claim 8, wherein said plurality of lists form a linked list.

10. The system of claim 8, wherein said plurality of lists is between four and eight.

11. The system of claim 8, wherein said computer system is a multitasking or multiprocessing computer system.

12. A method of operating a computer system having at least one processor, comprising: determining the validity of data in a memory unit; locking a list corresponding to said memory unit if said data is invalid; inserting an entry corresponding to one of said plurality of tasks onto said locked list; unlocking said locked list; and determining if data is inputted in said memory location between said determining step and said unlocking step.

13. A method for performing an operation within limits upon a shared value stored in an actual value location comprising: providing a plurality of memory locations, two of said memory locations being first and second reservation memory locations and two of said memory locations being limit memory locations containing limit values; getting an operand; performing the operation upon an affected reservation register using the operand; comparing a resulting value of the first operation upon the affected reservation register to the limit values in the limit memory locations; if the resulting value is not within the values in the limit memory locations, then performing an inverse operation to restore the affected reservation register and reporting a failure; and if the resulting value is within the values in the limit memory locations, then performing the operation to update the actual value location, performing the operation to update an unaffected reservation register, and report a success.

14. A system for performing an operation within limits upon a shared value stored in an actual value location comprising: a plurality of memory locations, two of said memory locations being first and second reservation memory locations, and two of said memory locations being limit memory locations containing limit values.