TWI489393B - Applied Assignment Method for Multi - core System - Google Patents

Description

Work distribution method applied to multi-core systems

The present invention relates to a multi-core system; in particular, to a work distribution method applied to a multi-core system.

Due to the limitation of semiconductor manufacturing process, physical characteristics and heat dissipation technology, the computing speed of the processor is unlikely to have a significant breakthrough in the short term. At present, the general response of the information industry is to use multiple processors (commonly referred to in the market). The concept of "core" integrated into a single device has been widely used in a variety of electronic products ranging from smart phones to high-end servers. In the case of a Chip Multi-Processor (CMP) product, multiple processors are included, so multiplex processing can be performed. Most systems with multiple processors generally use the last-level cache. (Last Level Cache, LLC) technology, which allows multiple processors to share a cache to avoid resource idleness and speed up system performance. However, since each task (Task) has different requirements for accessing the cache during execution (the "work" described here, according to the definition of the field of knowledge, it refers to the inclusion of a set of associated actions, and A program that can perform a specific function. If the processors sharing the same cache access the cache in an excessively dense manner in order to perform their respective tasks, it is likely that a large number of cache misses (Cache Miss) will occur. That is, the processor can't find the data needed to perform the work in the cache, and can only take time to move the data from the memory, resulting in a decrease in the number of executions per cycle (Instructions Per Cycle, IPC), which hinders system performance. . Therefore, in a multi-core system, how to avoid multiple processors to share as fast as possible Taking resources to snatch it has become an important issue.

The most fundamental way to solve the above problems is to seek a better way to distribute work to each processor. The goal is to enable multiple processors sharing the same cache to perform cache operations for each job. The access requirements can be met as much as possible. One of the conventional methods of work distribution is to use the Working Set Size (WSS) as the basis for distribution, so that the total size of the working set of all the work that is shared by multiple processors sharing the same cache is not Exceeding the cache capacity can also reduce the possibility of cache misses; another method is to distribute the work with higher misses per instruction (Misses Per Instruction (MPI) to different caches. The implementation of the device to evenly distribute the adverse effects of these tasks on system performance, but neither of these methods can truly reflect the degree of demand for access to the cache, so the effect is not good.

In addition, any work that is performed by the processor must have access to the cache used by the processor, so that it is somewhat interfered with by other jobs that have access requirements for the same cache. The ability to resist different work disturbances is not the same, so the ability to resist other work disturbances can be regarded as a nature of work, the so-called Anti-interference Ability. In short, the harder the work, the less likely it is to be affected by other work, so the shorter the execution time that is affected and stretched. At present, the conventional work distribution method is also based on the work immunity, such as the work with strong immunity and the work with weak immunity are respectively allocated to different processors sharing the same cache. Intuitively, this should reduce the phenomenon of resource grabbing; but in fact, the strength of the work is not related to the ability of the work to interfere with other work (called Interference Ability); It is said that work with strong anti-interference is not easy to interfere with other work, this method does not consider this. Although there are also simultaneous considerations Immunity and disturbances, but the need to use special binary testing tools, can not be immediately applied to a variety of systems.

Of course, there are many scholars who propose different methods, such as assigning work with high cache miss rate and work with low cache miss rate to processors sharing the same cache, or avoiding the same Work with low immunity is assigned to processors that share the same cache. However, the former needs to have the same number of jobs and the number of processors, while the latter requires the system to use First In First Out (FIFO) scheduling to process the work, in order to get better results, but the multi-core system This is not the case with the actual operation, so these two methods can only be used as preliminary research in the laboratory and cannot be applied in practice.

In view of this, the object of the present invention is to provide a work distribution method applied to a multi-core system, which can effectively reduce resource robbing of a shared cache by multiple processors, and the effect thereof is superior to the conventional methods. The scheduling of the system is limited and can handle a lot of work than the processor.

In order to achieve the above object, the present invention provides a work distribution method for a multi-core system, the multi-core system including a plurality of caches and a plurality of processor groups, each of the processor groups including at least two a processor, each of the caches being commonly used by the processors in the processor group; the work distribution method is for allocating a plurality of tasks to the processes in each of the processor groups Executing, and each processor can accept more than one assignment of work; each of the tasks defined herein has two properties: anti-interference ability and interference potential, and any of the tasks The immunity is defined as the work resists the same process when the work is performed by a processor within the processor group. The ability of other processors within the cluster to perform an impact on their execution time, any of which is defined as the operation being performed by a processor within a processor group. The work affects the execution time of other work performed by other processors within the same processor group; the work distribution method includes the following steps: A. Assessing the immunity and the disturbance of each work, and simultaneously Based on the immunity and the disturbance of each work, the work is divided into a plurality of work groups, and the number of the work groups is equal to the total number of the processors in the processor groups; Each of the working groups is allocated to each of the processors in each of the processor groups one by one until each of the processors in the processor group is assigned a working group.

The present invention further provides another work distribution method for a multi-core system, the multi-core system includes a plurality of caches and a plurality of processor groups, each of the processor groups including at least two processors, each The cache is used by a plurality of processors in the processor group; the work distribution method is configured to allocate a plurality of tasks to the processors in each of the processor groups, and Each of the processors can accept more than one assignment of work; each of the tasks defined herein has two properties, an anti-interference ability and an interference capability, and the immunity of any of the operations is defined. Defined as the ability of the work to resist the effects of other work performed by other processors within the same processor group on its execution time when executed by a processor within the processor group, either The scrambling force of the job is defined as the execution time of the other work performed by other processors within the same processor group when the work is performed by a processor within the processor group. The work distribution method includes the following steps: A. Assessing the immunity of each work, each of the work and an offensive work are respectively performed by a processor in the processor group, and the measurement is performed. Each of the first execution time required for the execution of the work is performed, and each of the operations is separately performed by a processor in the processor group. a second execution time required; for each of the work, the first execution time and the second execution time have a difference, and the ratio of the difference to the second execution time is an resistance of each work Disturbance, wherein the offensive work is a work that actively robs the cache resources when it is executed; B. Evaluate the disruptive power of each work, and each of the work and a defense work are assigned to the processor group. The second processor executes, measures a third execution time required for each work to be completed, and measures the defense work to be performed by a processor in the processor group. a fourth execution time; for each of the work, the third execution time and the fourth execution time have a difference, and the ratio of the difference to the fourth execution time is the disturbance of each work The defense work is a work that will regularly and stably use the cache resources when executed; C, respectively, the disturbance immunity and the disturbance power of each work are defined as a slowing factor of each work, The work is sorted in descending order of the respective slowing factors And sequentially dividing into a plurality of working groups, and the number of the working groups is equal to the total number of the processors in the processor groups; D, assigning each working group to each of the processor groups one by one Each of the processors until each of the processors in the processor group is assigned a working group.

Thereby, the present invention can enable multiple processors in the multi-core system to reduce the resources for robbing the shared cache when performing work, thereby improving system performance.

1 is a flow chart of a preferred embodiment of the present invention; FIG. 2 is an exemplary schematic view of a practical application of the present invention; and FIG. 3 is another exemplary schematic diagram of a practical application of the present invention.

In order to explain the present invention more clearly, the preferred embodiment is described in detail with reference to the accompanying drawings. The multi-core system to which the present invention is applied to the work distribution method of the multi-core system includes a plurality of caches (Cache). And a plurality of processor groups, each of the caches corresponding to a processor group, and each of the processor groups includes at least two processors, that is, each of the caches is included in a processor group These processors are used together and are described first.

The present invention is directed to assigning a plurality of tasks to the processors within each of the processor groups, and each of the processors can be assigned more than one job. Referring to FIG. 1, the present invention is based on evaluating the two properties of the anti-interference ability and the interference capability of each of the tasks. Determining the immunity of any of the operations herein means that the operation is performed by another processor within the same processor group when executed by a processor within the processor group. The ability of other work to affect its execution time; and the disturbing force of any such work means that the work affects the same process when it is executed by a processor within a processor group. The ability of other processors within the cluster to perform other work execution times. In short, the more robust the work, the less likely its execution time will be affected by the execution of other work; the more disruptive the work, the easier it will be to affect the execution time of other work.

In the preferred embodiment, the method for evaluating the immunity of each of the operations is performed by respectively assigning each of the work and an offensive work to a processor in the processor group, and measuring the execution of each work. Determining a first execution time, separately measuring a second execution time required for each of the work to be performed by a processor in the processing group; in other words, each of the work required to complete the work The second execution time, that is, the execution of each of the work is performed by a processor without being interfered by other work and the offensive work. Time required.

The offensive work is a work of actively absorbing resources. The virtual code shown in Table 1 is a demonstration of the offensive work, but it is not limited to this. It can be seen from the second to seventh rows of the virtual code that the offensive work performs a plurality of loops, and each action performs two actions in each loop, firstly generating three irregular random values x ₁ , x in random numbers. ₂ and x ₃ (lines 3 to 5), with reference to these three random values, in a two-dimensional array (Array) B , taken out at coordinates [ x ₂ ][ x ₃ ] and [ x ₁ ][ After the values of x ₃ ] are added together, they are stored in the coordinates [ x ₁ ][ x ₂ ] of the two-dimensional array B (line 6); since the values of x ₁ , x ₂ and x ₃ are all random numbers, The value generated in each loop may vary greatly, and please refer to the first line of the virtual code in Table 1. It can be seen that the size of the two-dimensional array B used in the offensive work is 400MB, far exceeding the current common cache capacity. ???the processor responsible for performing the offensive work must continuously store the data from the memory into the cache used by the processor, causing a work performed by another processor belonging to the same processor group. Execution time is dragged down. Therefore, for each work, the first execution time must be longer than the second execution time, and there is a difference between the two, and the ratio of the difference to the second execution time is defined as each Immunity to work.

Briefly, in the preferred embodiment, the immunity of each of the operations represents that the work is performed by a processor in the processor group. In the case of a line, the required execution time is stretched by the impact of the offensive work performed by another processor within the same processor group.

In the preferred embodiment, the method for evaluating the disturbance of each work is performed by each of the two processors in the processor group, and each of the work is performed. a third execution time required, and measuring a fourth execution time required for the defense work to be performed by a single processor; in other words, the fourth execution time required for the defense work to be completed, That is, the time required for the defense work to be completed by a processor without being disturbed by other work.

The defense work is a work that regularly and stably uses the cache resource when executed. The virtual code shown in Table 2 is a demonstration of the defense work, but is not limited thereto.

It can be seen from the 2nd to 9th rows of the virtual code of Table 2 that the defense work performs a plurality of loops, and the element value of a two-dimensional array B is modified according to a simple rule in each loop, and the first row of the virtual code of Table 2 The two-dimensional array B used for the defense work is initialized to a 2M size, which in the preferred embodiment is exactly equal to the capacity of the cache. Therefore, when the processor performs the defense work, if other processors belonging to the same processor group do not perform other work, there is no need to move the data from the memory into the cache used by the processor. The fourth execution time is the minimum time required for the theoretical implementation of the defense work. Therefore, for each of the work, the third execution time must be longer than the fourth execution time, and there is a difference between the two, and the ratio of the difference to the fourth execution time is defined as each The work is disturbing.

Briefly, in the preferred embodiment, the scrambling force of each of the operations represents that each of the tasks is performed by a processor within the processor group, and is within the same processor group. The execution time required for the defensive work performed by the other processor affects the proportion of execution time that is stretched.

After evaluating the immunity and the disturbance of each work, the present invention divides the work into a plurality of work groups according to the immunity and the disturbance of each work, and the number of the work groups is equal to The number of all processors in this multi-core system. For averaging, the number of jobs included in any two work groups differs by no more than one; as to how many jobs should be included in each work group, depending on the total number of jobs and the expected share of the work Depending on the number of groups, this is a simple mathematical operation and will not be described here.

The number of these work groups must be the same as the number of all processors of the multi-core system, because the work distribution method applied to the multi-core system of the present invention will eventually allocate the work groups one-to-one to each of the work groups. The processor; in other words, after the allocation is completed, each of the processors can be responsible for performing the work in the working group. The detailed allocation steps and descriptions are described later.

The classification basis used in the preferred embodiment is to add up the interference immunity and the disturbance power of each of the work, respectively, and define it as a slowing down factor for each work. Please refer to FIG. 2 , which is an application example of the preferred embodiment. In FIG. 2( a ), there are seven jobs A to G to be allocated, and it is assumed that the anti-interference force and the disturbance power of each work are evaluated by the foregoing steps, and After summing up the slowing down factors of the work, the work is sorted according to the decreasing factor of the slowing factor to obtain the order E, G, A, D, B, F, C as shown in Fig. 2(b). The slowing factor for this work is left As shown in FIG. 2(d), the multi-core system to which the preferred embodiment is applied has two caches, each of which is commonly used by two processors (ie, a common use) The cached processor can be regarded as a processor group, so the multi-core system has a total of four processors, and the work depends on the number of the processors, according to the order of the slowing factors of the work. Divided into four working groups (E, G), (A, D), (B, F), (C) as shown in Figure 2 (c), as can be seen from the figure, the more left working group, including The work has a greater slowing factor.

The work distribution method of the present invention applied to a multi-core system is prioritized by a work group including work having a large slowdown factor, and each of the work groups is assigned to each of the processors of the processor group one by one. Taking Figure 2 as an example, the first work group to accept the assignment is the leftmost work group (E, G) in Figure 2(c), followed by the work group (A, D), then the work group. (B, F), and finally the working group (C).

The method of assigning each of the work groups to each of the processors in the processing group has substantially two steps. In order to obtain the most efficient utilization of each of the multi-core systems, each of the working groups is first allocated to each processor in the processor group in turn until there is one in each of the processor groups. The processor is assigned a working group. Taking FIG. 2 as an example, as shown in FIG. 2(d), the working group (E, G) is first allocated to a processor in the processor group located on the left side, and then the working group (A, D) is A processor is allocated to the processor group located on the right side, such that each of the processor groups in the multi-core system has a processor assigned the work. It should be noted that although the examples herein are allocated to each processor group from left to right, since the performance of each processor group is virtually the same, this directionality is not a limitation of the present invention, and other implementations. In the example, of course, from right to left, or even in another multi-core system with more processor groups, there may be more varied allocation methods, as long as the process of assigning the working groups in sequence is ensured. , the There will be one in the processor group to allocate the working group.

Accordingly, the processor groups necessarily have an assigned sequence (as from left to right in the above example), which is defined herein as an allocation order. The work groups that have not been allocated are preferentially allocated to a processor group located at the end of the allocation sequence. After all the processors in the processor group are assigned the work group, the remaining work groups are reconciled. The processors in the other processor groups are reversely allocated along the allocation order until all processors owned by the multi-core system are assigned a working group. Since the number of the work groups is the same as the total number of the processors of the core system, the work groups have been assigned to the processor, that is, the work in each work group has been The allocation is completed and assigned to a processor for execution.

Also taking the example of FIG. 2 as an example, the allocation order is from left to right as described above, and after each processor group has one of the working groups assigned to the working group, the remaining working groups (B, F) And (C) proceeding in the reverse order of the allocation, preferentially assigning the working group (B, F) to other processors in the processor group located on the right side (in this example, only the remaining one in the processor group) Processor); finally, the work group (C) is assigned to the last remaining processor in the processor group on the left side, and all the work is then assigned, and the final result is shown in Figure 2(d).

Please refer to FIG. 3, which is another application example of the preferred embodiment. In FIG. 3(a), there are fourteen to four work to be assigned, and it is assumed that the immunity and the disturbance of each work are evaluated, and After summing up the slowing down factors for each of the jobs, the jobs are sorted by left and right according to their slowing factors into the order E, I, G, A, B, F, as shown in Figure 3(b). C, J, D, H, N, L, M, K; see FIG. 3(d), the multi-core system to which the preferred embodiment is applied has two caches, each of which is composed of four processors Used together (that is, the processors that use the cache together are regarded as a processor group), so the multi-core system has a total of eight processors, so the work is based on each work. The slowing factor is sorted into eight working groups (E, I), (G, A), (B, F), (C, J), (D, H) as shown in Fig. 3(c). , (N, L), (M), (K), the work group on the left side of the figure, the greater the slowing factor of the work involved.

According to the foregoing allocation step, the allocation is started from the leftmost side in FIG. 3(c), which is the same as the above example provided in FIG. 2, and the working groups (E, I) and (G, A are also taken from the left-to-right allocation order. Each of the processors is allocated to each of the processor groups, but the order of allocation is not a limitation of the present invention; in this case, each processor group of the multi-core system has a processor assigned to the working group. Next, in the reverse of the allocation order, the working groups (B, F), (C, J), (D, H) are respectively allocated to the processors in the processor group located on the right side that have not been allocated; Finally, the work groups (N, L), (M), (K) are respectively assigned to the remaining processors in the processor group located on the left side. At this point, all work has been assigned to the processors in the processor groups and is executed by the processors.

It should be specially noted that in the examples of FIG. 2 and FIG. 3, the number and arrangement order of each work (and the grouping result affected by the arrangement order) are purely for demonstration purposes, just for convenience of explanation and understanding. It does not have any special meaning.

In summary, the working distribution method of the present invention applied to a multi-core system can reduce the robbing of cache resources by multiple processors in a multi-core system, and has positive effects on system performance improvement; The second cache is tested on a multi-core system. Compared with the conventional scheduling method known in the Linux operating system, the present invention saves 43% of the total execution time, and the number of execution instructions per cycle (IPC) is increased by 51%.

The above description is only for the preferred embodiments of the present invention, and the equivalent method variations of the present invention and the scope of the patent application are intended to be included in the scope of the present invention.

Claims (12)

A work distribution method for a multi-core system, the multi-core system includes a plurality of caches and a plurality of processor groups, each of the processor groups including at least two processors, each of the caches being The processor in the processor group is used in common; the work distribution method is configured to allocate a plurality of tasks to the processors in each of the processor groups, and each of the processors can be Accepting the assignment of more than one work; each of the tasks defined here has two properties: anti-interference ability and interference potential. The immunity of any of the tasks is defined as the work. When performed by a processor within the processor group, the work resists the ability of other operations performed by other processors within the same processor group to affect its execution time, either of which interferes with the operation. Force is defined as the ability of the work to affect the execution time of other jobs performed by other processors within the same processor group when executed by a processor within the processor group; The method comprises the following steps: A. Assessing the immunity and the disturbance of each work, and dividing the work into a plurality of work groups based on the immunity and the disturbance of each work, and the The number of working groups is equal to the sum of the processors in the processor groups; B. assigning each working group to each processor in each processor group until each processor group Each of the processors is assigned a working group. The method for assigning a work to a multi-core system as described in claim 1, wherein the method for evaluating the immunity of each of the work in step A is to respectively assign each of the work and an offensive work to a processor group. The processor executes, measures a first execution time required for each execution of the work, and measures each of the operations required to be performed by a processor in the processor group. a second execution time; for each of the work, the first execution time and the second execution time have a difference, and the ratio of the difference to the second execution time is the immunity of each work; Among them, the offensive work is a work that actively robs the cache resources when it is executed. The work distribution method applied to the multi-core system according to claim 2, wherein the offensive work performs a plurality of loops, and generates a plurality of random values in each loop, and then modifies according to the random values. An element value of a multidimensional array (Array); wherein the size of the multidimensional array is much larger than the size of the cache used by the processor performing the offensive work. The method for assigning a work to a multi-core system as described in claim 1, wherein the method for evaluating the disturbance of each work in step A is to respectively assign each work and a defense work to a processor group. The processor executes, measures a third execution time required for each work to be completed, and measures the defense work to be performed by a processor in the processor group. a fourth execution time; for each of the work, the third execution time and the fourth execution time have a difference, and the ratio of the difference to the fourth execution time is the disturbance of each work; The defense work is a work that regularly and stably uses the cache resources when executed. The work distribution method applied to the multi-core system according to claim 4, wherein the defense work performs a plurality of loops, and the element values of a multi-dimensional array are successively modified in a simple rule in each loop; wherein the multi-dimensional The size of the array is exactly the size of the cache used by the processor performing the defense work. The work distribution method applied to the multi-core system as described in claim 1, wherein the difference in the number of the jobs included in any one of the work groups described in step A is at most one. The method for applying the work to the multi-core system as described in claim 1, wherein the method for allocating each of the work groups to each of the processors in the processor group in the step B further includes the following steps: And each of the working groups is allocated to each processor in the processor group in turn until each processor group has a working group assigned to the processor group, wherein the processor groups accept the order of allocation Defined as an allocation order; B2, reversed along the allocation order, and the remaining working groups are sequentially allocated to all the processors in the processor group that have not been assigned the working group until the processing The processors in the cluster are all assigned a working group. A work distribution method for a multi-core system, the multi-core system includes a plurality of caches and a plurality of processor groups, each of the processor groups including at least two processors, each of the caches being A processor in the processor group is used in common; the work distribution method is used to allocate a plurality of tasks to the processors in each of the processor groups. And each processor can accept more than one assignment of work; each of the tasks defined herein has two properties: anti-interference ability and interference potential, and any of the work is immune to interference. Force is defined as the ability of the work to resist the effects of other work performed by other processors within the same processor group on its execution time when executed by a processor within the processor group. A disturbing force of the work is defined as the execution time of the other work performed by other processors within the same processor group when the work is performed by a processor within the processor group. The work distribution method includes the following steps: A. Assessing the immunity of each work, each of the work and an offensive work are respectively performed by a processor in the processor group, and the measurement is performed. a first execution time required for each work to be completed, and measuring a second execution time required for each of the work to be performed by a processor in the processor group; The first execution time and the second execution time have a difference, and the ratio of the difference to the second execution time is an immunity of each work, wherein the offensive work is performed when executed. Actively robbing the work of caching resources; B. Evaluating the disruptive power of each of the tasks, respectively, each of the work and a defense work are performed by a processor in the processor group, and the measurement is obtained. a third execution time required for the execution of the work to be completed, and measuring a fourth execution time required for the defense work to be performed by a processor in the processor group; for each work The third execution time and the fourth execution time have a difference, the difference and the fourth execution time The ratio between the two is the disturbance of each work; the defense work is a work that will regularly and stably use the cache resources when executed; C, respectively, the disturbance immunity and the disturbance power of each work are added, Defining a slowing factor for each of the work, sorting the work according to the respective slowing factors, and sequentially dividing into a plurality of working groups, and the number of the working groups and the number of the working groups The sum of the processors is equal; D. Each work group is allocated to each processor in each processor group one by one until each processor in each processor group is assigned a working group. The work distribution method applied to the multi-core system according to claim 8, wherein the attacking work described in step A performs a plurality of loops, and generates a plurality of random values in each loop, and then The random values modify the element values of a multidimensional array (Array); wherein the size of the multidimensional array is much larger than the size of the cache used by the processor performing the offensive work. The work distribution method applied to the multi-core system as described in claim 8, wherein the defense work described in step B performs a plurality of loops, and the elements of a multi-dimensional array are successively modified according to a simple rule in each loop. a value; wherein the multidimensional array is exactly the size of the cache used by the processor performing the defense work. The work distribution method applied to the multi-core system as described in claim 8, wherein the difference in the number of the jobs included in any one of the work groups described in step C is at most one. The method for allocating the work to the multi-core system as described in claim 8, wherein the method for allocating each of the work groups to each of the processors in the processor group further includes the following steps: D1 And each of the working groups is allocated to each processor in the processor group in turn until each processor group has a working group assigned to the processor group, wherein the processor groups accept the order of allocation Defined as an allocation order; D2, reversed along the allocation order, and the remaining working groups are sequentially assigned to all the processors in the processor group that have not been assigned the working group until the processing The processors in the cluster are all assigned a working group.