TWI768554B - Computing system and performance adjustment method thereof - Google Patents

Abstract

A computing system and a performance adjustment method thereof are provided. In the method, a performance test is performed according to a performance arrangement. The computing system includes two or more processors. The performance test is used for a load scenario, and the performance arrangement is related to the weights of the processors to make efforts on the performance under the load scenario. A gradient corresponding to the test result of the performance test is determined. The performance test is related to a multi-variant function that is used to present the performance of the performance test. The gradient is obtained by performing gradient calculation on the multi-variant function with the performance arrangement and the test result. The performance arrangement is modified according to the gradient. Accordingly, better overall performance under the system limitation can be obtained.

Description

Computing system and method for adjusting performance thereof

The present invention relates to a multiprocessor performance tuning technology, and more particularly, to a computing system for multiprocessors and a performance tuning method thereof.

In order to meet the application in various fields, electronic products today are gradually adopting Heterogeneous Computing system design from High Performance Computing (HPC) servers, personal computers, and even smart phones to achieve better performance. Calculate benefit or energy efficiency. For example, an HPC server is equipped with a central processing unit (CPU), a graphics processing unit (GPU), a field programmable gate array (FPGA), and a special application integrated circuit. (Application-Specific Integrated Circuit, ASIC), special specification accelerator card and/or neural network accelerator. In addition to the original independent CPU combined with GPU design in personal computers and mobile phones, the CPU design has also begun to move toward the trend of integrating cores of different computing capabilities or functions into a single chip, and adopting more sophisticated features for different workloads The dominant computing unit achieves the best computing efficiency.

A common design problem for multiple computing units is that the computing power of each computing unit is tied to each other by the system design. For example, CPU and GPU share thermal capacity due to the connection of heat pipes. If one computing unit runs at high speed, the computing power of the other computing unit will be squeezed. Another common design problem is that the design of products with insufficient wattage will cause the modules that perform operations first in time to seize the power budget, thereby compressing the performance of other modules. Both of these common cases reduce the advantage of heterogeneous computing.

In view of this, embodiments of the present invention provide a computing system and a performance adjustment method thereof, which adjust the performance configuration of multiprocessors (ie, computing units) based on gradient characteristics, so as to quickly obtain a better configuration under system constraints.

The performance adjustment method of the computing system according to the embodiment of the present invention includes (but is not limited to) the following steps: performing a performance test according to the performance configuration. The computing system includes two or more processors. The performance test is for a load situation, and the performance profile is related to the proportion of those processors that pay performance under the load situation. Determines the gradient corresponding to the test results of the performance test. Test results are related to a multivariate function used in performance testing to determine performance. This gradient is obtained by performing gradient operations on a multivariate function based on that performance configuration and test results. Modify the performance configuration according to this gradient.

The computing system of the embodiment of the present invention includes (but is not limited to) two or more processors. The processors are configured to perform a performance test according to the performance configuration, determine a gradient corresponding to the test results of the performance test, and modify the performance configuration according to the gradient. The performance test is for a load situation, and the performance profile is related to the proportion of those processors that pay performance under the load situation. Test results are related to a multivariate function used in performance testing to determine performance. This gradient is obtained by performing gradient operations on a multivariate function based on that performance configuration and test results.

Based on the above, in the computing system and the performance adjustment method thereof according to the embodiments of the present invention, a gradient operation is performed on the multivariate performance function under a specific load situation (ie, workload) to understand the trend of performance growth, and then adjust those processors performance configuration. Thereby, better overall performance can be achieved.

In order to make the above-mentioned features and advantages of the present invention more obvious and easy to understand, the following embodiments are given and described in detail with the accompanying drawings as follows.

FIG. 1 is a block diagram of a computing system 100 according to an embodiment of the invention. Referring to FIG. 1 , the computing system 100 may be a desktop computer, a notebook computer, an AIO computer, a smart phone, a tablet computer, or a server. Computing system 100 includes, but is not limited to, memory 110 and two or more processors 130 .

The memory 110 may be any type of fixed or removable random access memory (RAM), read only memory (ROM), flash memory, conventional hard disks (Hard Disk Drive, HDD), Solid-State Drive (Solid-State Drive, SSD) or similar components. In one embodiment, the memory 110 is used to record code, software modules, configuration, data or files.

The processor 130 is coupled to the memory 110. The processor 130 can be a CPU, a GPU, or other programmable general-purpose or special-purpose microprocessors (Microprocessors), digital signal processors (DSPs) , programmable controllers, FPGAs, ASICs, neural network accelerators or other similar elements or combinations of the above. In one embodiment, the processor 130 is used to execute all or part of the operations of the computing system 100 , and can load and execute various software modules, files and data recorded in the memory 110 .

In one embodiment, those processors 130 may be of different types if a heterogeneous computing system is to be implemented. For example, computing system 100 includes a CPU and a GPU. In some embodiments, some or all of the processors 130 may be of the same type. For example, computing system 100 includes three GPUs.

Hereinafter, the methods described in the embodiments of the present invention will be described in conjunction with various devices, components and modules in the computing system 100 . Each process of the method can be adjusted according to the implementation situation, and is not limited to this.

FIG. 2 is a flowchart of a performance adjustment method according to an embodiment of the present invention. Referring to FIG. 2, the processor 130 may perform a performance test according to the performance configuration (step S210). Specifically, the performance test is, for example, Cinebench, 3DMark, 3DMark11, AnTuTu, or other types of performance or benchmark tests for evaluating load, stress, breakpoints, and the like. This benchmark is for a specific load scenario (eg, time spy, cloud processing of servers, neural network training/inference, video playback, gaming, video editing, web browsing, video chat, floating point arithmetic, etc.). In addition, the performance profile is related to the proportion of those processors 130 that pay performance under this load situation. In other words, the proportions represent the proportions that affect the overall performance when the processors 130 work together. The performance profile can be changed by changing factors such as the execution frequency, voltage, power, or fan speed of the processor 130 . For example, the load situation of 3D model editing needs to set a higher weight on the GPU.

The processor 130 may determine a gradient corresponding to the test result of the performance test (step S230). Specifically, test results are related to a multivariate function used in performance testing to determine performance. This multivariate function is related to the performance of those processors 130 and the overall performance of the computing system 100 under this load situation. For example, performance as a variable and overall performance as a function value. In some embodiments, the computing system 100 may obtain test results from other computing systems 100 .

…(1) 其中

為各處理器130在當前負載情境下的個別測試分數(或個別測試表現)，

為各處理器130在此負載情境下的權重，n為處理器130的個數，i為正整數。 In one embodiment, the test results include individual test scores for those processors 130 and overall test scores. The individual test scores of each processor 130 are used as variables, and the overall test scores are used as the solution of the function. That is, the overall test score is obtained by inputting the individual test scores of the processor 130 into a multivariate function. The multivariate function takes the Weighted Harmonic Mean performance function P as an example:

…(1) of which

is the individual test score (or individual test performance) of each processor 130 under the current load situation,

is the weight of each processor 130 under this load situation, n is the number of processors 130, and i is a positive integer.

It should be noted that the multivariate functions of different load situations may be different, which is not limited in the embodiment of the present invention.

Gradients, on the other hand, are obtained by performing gradient operations on multivariable functions based on performance configurations and test results. In vector calculus, a gradient is a generalization about multivariate/multivariate derivatives. The gradient of a multivariate function is a vector-valued function. The gradient of a multivariable differentiable function at a point on a multidimensional vector space is a vector whose component is the partial derivative of the multivariable differentiable function at this point. For a multivariate function of performance, the direction of this vector is the maximum increase in performance for this function at that point, and its components represent the growth rate of performance in this direction. From this, it can be seen that gradient-based characteristics can learn how to increase the function value (ie, overall performance) of a multivariate function for a specific load situation.

…(2) 其中

為這些處理器130在所形成的n維向量空間中使處理器130的效能增加方向為正的單位向量，

為第i處理器130對應向量的分量大小。 In one embodiment, the processor 130 may determine the gradient vectors in the gradients that correspond to those of the processor 130 . This gradient vector includes magnitudes and directions of components corresponding to those of the processor 130 vector, respectively. Taking function (1) as an example, the gradient operation on function (1) can be obtained:

…(2) of which

For these processors 130 in the formed n-dimensional vector space, the direction of the performance increase of the processors 130 is a positive unit vector,

is the component size of the vector corresponding to the i-th processor 130 .

…(3) 其中

為GPU和CPU在當前負載情境下的個別測試分數，0.85:0.15為GPU和CPU在此負載情境下的權重。 Taking the load situation of 3Dmark Time Spy as an example, the multi-variable function of its performance is:

…(3) of which

are the individual test scores of GPU and CPU under the current load situation, and 0.85:0.15 is the weight of GPU and CPU under this load situation.

A multivariate function can be represented as a multidimensional surface in a multidimensional vector space. Taking function (3) as an example, FIG. 3A is an example illustrating the performance surface. Referring to FIG. 3A, given a CPU and GPU individual test scores (first score, second score), the overall performance performance (first score, second score) coordinate upwards corresponding to the overall score of the point on the surface is obtained Coordinates, where the direction of the arrow is defined as the direction of increased performance.

The coordinates of each point represent a possible setting rule. For example, the point T1 shown in FIG. 3A represents the result of a certain performance test. In this test, the ratio of performance configurations that limit both GPU and CPU is about 1.3:1. In the same computing system 100, the performance of the GPU is greatly relaxed from the position of the point T1 shown in FIG. 3A, but the performance of the CPU is limited. Assume that the individual test scores S1 and S2 of GPU and CPU are 7,354 and 5,657 respectively, and the overall test score S3 is 7,037.

FIG. 3B is another example illustrating the performance surface. Please refer to FIG. 3B, assuming that the ratio of GPU and CPU performance configuration is about 5:1 to obtain this test result. For this configuration modification (located at point T2), although the GPU's individual test score S1 was greatly improved (from 7,354 to 10,041), the overall test score S3 dropped from 7,037 to 6,437.

It is also worth noting that, a common system design method in the present is to set corresponding rules for each processor 130 given the results obtained by experience or specific Benchmark repeated experiments. For example, the load situation of 3DMark Time Spy uses a lot of DirectX 12 API for image calculation, so the general direction of configuration optimization is to limit the wattage used by the CPU but allow the thermal capacity to be used by the GPU. However, in practice, how to set the optimal balance point, the existing trial and error fine-tuning has no systematic rules to follow, and it is also quite time-consuming.

In the embodiment of the present invention, the processor 130 may modify the performance configuration according to the gradient (step S250). Specifically, the embodiment of the present invention can bring the operation result (ie, the test result) under a certain setting rule (ie, the performance configuration) into the gradient function (eg, function (2)), by increasing the gradient function with The capabilities of the processor 130 with the largest vector component while reducing the capabilities of other processors 130 to achieve better overall performance (eg, overall test scores).

…(5) For example, the function (4) can be obtained after the gradient operation on the function (3):

…(5)

It is worth noting that, as can be seen from the aforementioned examples of FIGS. 3A and 3B , for a specific load situation, based on the known test results of FIG. 3A , the performance configuration of FIG. 3B should not be continued, and such configuration changes will affect the overall performance. Efficacy doesn't help. And if each point on the multi-dimensional surface is to get the highest vertical axis rise (ie, the overall test score increases), the performance weight of GPU and CPU can be determined by a gradient function (eg, function (4)).

FIG. 4 is an example illustrating gradient vectors. Please refer to FIG. 4 , the gradient vector field of function (4) is a two-dimensional vector field, which is to project the vector field on the horizontal plane where the value of the vertical axis is zero as shown in FIG. 3A or FIG. 3B , and is normalized Figure 4 can then be obtained. The direction of the arrow (ie, the direction of the vector) at each location represents how this performance configuration should be configured locally to achieve the maximum overall test score improvement.

和

的分量大小的比例，這樣可以提升最大的總體效能表現。圖3A中點T1的位置的梯度向量大約為

，且其正規化向量(

)大約為

。此正規化向量大約對應到圖4所示點T1。此配置下，向量

的分量大於

的分量，因此梯度向量的向量方向指向為左上，但較偏上。而基於梯度所得出的效能配置的比重為0.958:0.286 ≈ 3.3:1。處理器130可依據此比重修改原效能配置。例如，直接將此比重設定為新效能配置。又例如，適當增加部分比重。 In one embodiment, the processors 130 may modify the performance configuration according to the component sizes corresponding to those processors 130 . In other words, the component size of the unit vector corresponding to each processor 130 can be used as a basis for adjusting the weight of the performance configuration. Taking function (4) as an example, the performance configuration of CPU and GPU can be set to the unit vector in the gradient vector

and

The proportion of the component size, which can improve the maximum overall performance. The gradient vector at the location of point T1 in Figure 3A is approximately

, and its normalized vector (

) is approximately

. This normalized vector corresponds approximately to point T1 shown in FIG. 4 . In this configuration, the vector

component is greater than

The component of , so the vector direction of the gradient vector points to the upper left, but it is more upward. The proportion of the performance configuration based on the gradient is 0.958:0.286 ≈ 3.3:1. The processor 130 can modify the original performance configuration according to the weight. For example, simply set this weight to the new performance profile. For another example, a part of the specific gravity is appropriately increased.

In one embodiment, the processor 130 may determine the vector directions of those gradient vectors of the processor 130 in the multi-dimensional vector space. This multi-dimensional vector space is formed by dimensions corresponding to the individual test scores of those processors 130 . Taking FIG. 4 as an example, the individual test scores S1 and S2 (ie, the first score and the second score) of the GPU and the CPU are respectively defined as two dimensions.

The processor 130 can modify the performance configuration according to the direction of the vector direction. For example, increasing or decreasing the weight of the corresponding dimension towards the direction. For example, if more overall test scores (overall performance presentation) are to be added for the point T1 in FIG. 4 , the proportion of the GPU corresponding to the first score should be further increased, but the proportion of the CPU corresponding to the second score can be maintained ( Corresponds to the direction of the vector pointing to the upper left but more upward).

，且其正規化向量(

)大約為

。此正規化向量大約對應到圖4所示點T2。此配置下，向量

的分量大於

的分量，因此梯度向量的向量方向指向為左上，但較偏左。若欲針對圖4中的點T2增加更多整體測試分數(整體效能呈現)，則例如是第一分數對應的GPU的比重應降低，但第二分數對應的CPU的比重應增加。而基於梯度所得出的效能配置的比重為0.245:0.969 ≈ 1:3.96。處理器130可依據此比重修改原效能配置。 For the example of FIG. 3B (overly relaxing GPU performance), the gradient vector at the position of point T2 in FIG. 3B is approximately

, and its normalized vector (

) is approximately

. This normalized vector corresponds approximately to point T2 shown in FIG. 4 . In this configuration, the vector

component is greater than

The component of , so the vector direction of the gradient vector points to the upper left, but more to the left. To add more overall test scores (overall performance presentation) for point T2 in FIG. 4 , for example, the proportion of GPU corresponding to the first score should be reduced, but the proportion of CPU corresponding to the second score should be increased. The proportion of the performance configuration based on the gradient is 0.245:0.969 ≈ 1:3.96. The processor 130 can modify the original performance configuration according to the weight.

FIG. 5 is a flow chart illustrating system limit decisions according to an embodiment of the present invention. Referring to FIG. 5 , the processor 130 can determine whether the gradient reaches the system limit (step S510 ). Specifically, the system limitations relate to the performance configurations allowed by the computing system 100 . For example, supply wattage, cooling capability, or other factors that affect performance. For another example, the component size corresponding to the gradient of the CPU is about 5, but the supply wattage required for the specific gravity to exceed 4 is greater than the upper limit that the computing system 100 can provide, that is, it is determined that the system limit is reached. If the gradient has not reached the system limit, the processor 130 may modify the performance configuration based on the gradient vector (step S530 ), execute the performance test again according to the modified performance configuration (for example, execute step S210 ), and execute the performance test again according to the test result of the performance test Determine the corresponding gradient (for example, perform step S230). For example, if the ratio of the performance configuration obtained by the aforementioned gradient is 1:3.96, the processor 130 modifies the performance configuration of the GPU and the CPU to 1:3.96, and performs the performance test under the same load situation again to obtain the test result, and Accordingly, the corresponding gradient is obtained. This is repeated (the value should be able to converge) until the performance adjustment has reached the system design limit, the processor 130 will disable the adjustment of the performance configuration (ie, no longer adjust the performance configuration), and can determine the final performance configuration accordingly (step S550). At this point, the final performance configuration should be the configuration that provides the best overall performance for this particular load situation.

In practical applications, the processor 130 can record the performance configuration based on the gradient and corresponding to different load situations. In addition, the processor 130 can detect the current operation behavior of the computing system 100 (eg, game, video, or web browsing, etc.), and set the performance configuration corresponding to the load situation accordingly. In this way, the computing system 100 can achieve better performance in specific applications.

To sum up, according to the computing system and the performance adjustment method of the present invention, an appropriate performance configuration can be obtained by using the gradient corresponding to the test result of the performance test for a specific load situation. In this way, you can quickly find the optimal configuration under system constraints.

Although the present invention has been disclosed above by the embodiments, it is not intended to limit the present invention. Anyone with ordinary knowledge in the technical field can make some changes and modifications without departing from the spirit and scope of the present invention. Therefore, The protection scope of the present invention shall be determined by the scope of the appended patent application.

100: Computing Systems 110: Memory 130: Processor S210~S250, S510~S550: Steps T1, T2: point S1, S2: Individual test scores S3: Overall Test Score

FIG. 1 is a block diagram of a computing system according to an embodiment of the present invention. FIG. 2 is a flowchart of a performance adjustment method according to an embodiment of the present invention. FIG. 3A is an example illustrating a performance surface. FIG. 3B is another example illustrating the performance surface. FIG. 4 is an example illustrating gradient vectors. FIG. 5 is a flow chart illustrating system limit decisions according to an embodiment of the present invention.

S210~S250: Steps

Claims (8)

A performance adjustment method of a computing system, wherein the computing system includes at least two processors, and the performance adjustment method includes: executing a performance test according to a performance configuration, wherein the performance test is for a load situation, and the performance configuration is related to The at least two processors pay a proportion of performance under the load situation; determine a gradient corresponding to a test result of the performance test, wherein the test result is related to a multivariate function used for the performance test to determine performance, And the gradient is obtained by performing gradient operation on the multivariate function based on the performance configuration and the test result; modifying the performance configuration according to the gradient; re-executing the performance test according to the modified performance configuration; and re-executing the performance The test result of the test determines the corresponding gradient. The performance adjustment method of claim 1, wherein the step of determining the gradient corresponding to the test result of the performance test comprises: determining a gradient vector in the gradient corresponding to the at least two processors, wherein the gradient vector includes component sizes of the vector of the at least two processors; and modifying the performance configuration according to the component sizes corresponding to the at least two processors. The performance adjustment method of claim 2, wherein the test result includes individual test scores of the at least two processors and an overall test score, and the overall test score is obtained by inputting the individual test scores of the at least two processors into the multivariable The steps of modifying the performance profile according to the gradient include: determining the vector direction of the gradient vector of the at least two processors in a multidimensional vector space, wherein the multidimensional vector space is formed by a plurality of dimensions corresponding to the individual test scores of the at least two processors; and according to the vector The orientation of the direction modifies the performance profile, where the weight of the corresponding dimension is increased or decreased toward that orientation. A performance adjustment method for a computing system, wherein the computing system includes at least two processors, and the performance adjustment method includes: executing a performance test according to a performance configuration, wherein the performance test is based on a load situation, and the performance configuration relative to the proportion of performance paid by the at least two processors under the load situation; determining a gradient corresponding to a test result of the performance test, wherein the test result is related to a multivariate function used for the performance test to determine performance performance, and The gradient is obtained by performing a gradient operation on the multivariate function based on the performance configuration and the test result; modifying the performance configuration according to the gradient; determining whether the gradient reaches a system limit, wherein the system limit is related to the computing system a permissible performance profile; and disabling adjustment for the performance profile for the gradient up to the system limit. A computing system comprising: at least two processors, one of which is configured to perform: executing a performance test according to a performance configuration, wherein the performance test is for a load situation and the performance configuration is related to the at least two processors in this load situation The proportion of performance paid in the environment; determine a gradient corresponding to the test result of the performance test, wherein the test result is related to a multivariate function used for the performance test to determine the performance performance, and the gradient is based on the performance configuration and the The multi-variable function of the test result is obtained by performing gradient operation; modifying the performance configuration according to the gradient; executing the performance test again according to the modified performance configuration; and determining the corresponding gradient according to the test result of the performance test executed again. The computing system of claim 5, wherein one of the at least two processors is further configured to perform: determining a gradient vector in the gradient corresponding to the at least two processors, wherein the gradient vector includes respectively corresponding to component sizes of the vector of the at least two processors; and modifying the performance configuration according to the component sizes corresponding to the at least two processors. The computing system of claim 6, wherein the test result includes an individual test score of the at least two processors and an overall test score, the overall test score being inputted into the multivariate function by the individual test scores of the at least two processors is obtained, and one of the at least two processors is further configured to perform: determining the vector direction of the gradient vector of the at least two processors in a multi-dimensional vector space, wherein the multi-dimensional vector space is defined by corresponding to the forming a plurality of dimensions of the individual test scores of at least two processors; and modifying the performance configuration according to the direction of the vector direction, wherein the weight of the corresponding dimension is increased or decreased toward the direction. A computing system comprising: at least two processors, one of which is configured to perform: executing a performance test according to a performance configuration, wherein the performance test is for a load situation and the performance configuration is related to the at least two processors The proportion of performance that the device pays under the load situation; determining a gradient corresponding to the test result of the performance test, wherein the test result is related to a multivariate function used for the performance test to determine performance, and the gradient is based on the The performance configuration and the multivariate function of the test result are obtained by performing gradient operations; modifying the performance configuration according to the gradient; determining whether the gradient reaches a system limit, wherein the system limit is related to the performance configuration allowed by the computing system; and Adjustments for the performance profile are disabled for the gradient up to the system limit.

Images