Unified performance modeling and adaptive changing method and device for diversified computing power
An adaptable and high-performance technology, applied in the field of heterogeneous computing, can solve the problems of difficulty in locating global performance bottlenecks, inefficiency, and poor performance, and achieve the effect of supporting rapid deployment, flexible and frequent changes
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0064] In a heterogeneous computing scenario, each computing node can contain multiple computing devices, and the following common situations exist between computing devices:
[0065] a) Device types are different, such as CPU, GPU, DPU and FPGA and other types of devices;
[0066] b) The same device (such as CPU) but belongs to different instruction set architecture (InstructionSetArchitecture, ISA), such as x86, Arm and RISC-V, etc.;
[0067] c) the same device and the same ISA (such as x86 CPU), but different manufacturers, such as CPUs produced by Intel and AMD;
[0068] d) The same device, the same ISA, and the same manufacturer (such as Intel CPU), but belong to different generations. For example, the CPU produced by Intel has generations such as SkyLake, KabyLake, and CascadeLake.
[0069] For ease of description, this embodiment is described as two computing devices A and B in a computing node. The relationship between the two computing devices in a specific embodimen...
Embodiment 2
[0085] Taking the background of Embodiment 1 as an example, instead of upgrading the system equipment in a coexistence mode, computing equipment B is used to replace computing equipment A. In this case, it is necessary to analyze the improvement of system performance before and after the computing equipment is moved in.
[0086]As described in the first embodiment above, the device type, ISA, manufacturer, model, generation, etc. of the computing device B are not exactly the same as those of the computing device A. Moreover, in this embodiment, replacing computing device A with computing device B involves problems such as moving in of the new computing device and moving out of the old computing device. For this scenario, according to the adaptive change mechanism proposed in this embodiment, its implementation scheme is as follows Figure 5 shown.
[0087] This embodiment also uses the CPI in Embodiment 1 as perf_metrics. When the device type of the hardware device layer cha...
Embodiment 3
[0094] In the actual production operation environment, the aggregation of perf_metrics is firstly aggregated within the computing node, and then aggregated between computing nodes, which is a hierarchical propagation process. This embodiment expands the situation of computing nodes in the above embodiments, from a single computing node to multiple computing nodes, and explains how to realize unified performance modeling of diversified computing power. The situation in computing nodes is the same as that of Embodiment 1 and implementation Example two.
[0095] Take the heterogeneous computing system between computing nodes as an example, such as Figure 6 As shown, in a heterogeneous parallel computing system, two computing nodes are included, the processor of one computing node is an Intel processor of x86 architecture (for example, Intel XeonGold processor), and the processor of the other computing node is Arm-based Cortex processors (such as ArmNeoverse N1 processors). Now...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More