Estimating performance of GPU application for different GPU-link performance ratio

Abstract

A computer-implemented method is provided for estimating the performance of a GPU application on a new computing machine having an increased GPU-link performance ratio relative to a current computing machine having a current GPU-link performance ratio. The method includes adding a delay to CPU-GPU communication on the current computing machine to simulate a delayed-communication environment on the current computing machine. The method further includes executing the target GPU application in the delayed-communication environment. The method also includes measuring the performance of the target GPU application in the delayed-communication environment. The method additionally includes estimating the performance of the new computing machine having the increased higher GPU-link performance ratio, based on the measured performance of the target GPU application in the delayed-communication environment.

Description

BACKGROUNDTechnical Field[0001]The present invention relates generally to information processing and, in particular, to estimating the performance of a Graphical Processing Unit (GPU) application for a different GPU-link performance ratio.Description of the Related Art[0002]Currently, a supercomputer node often includes a CPU and a GPU. On such computing machines, application execution mainly involves “GPU computation” and “CPU-GPU communication”. That is, before executing a kernel on a GPU, the data used by the kernel has to be transferred from CPU memory to GPU memory. After execution, the data produced by the kernel on the GPU is typically transferred back to the CPU memory. In the preceding scenario, the execution relates to “GPU computation”, while the transferring of data relates to “CPU-GPU communication”[0003]Performance estimation for future computing machines, especially for machines with faster GPUs, is very important both for application tuning and for sales. However, th...

AI Technical Summary

Benefits of technology

This approach allows for accurate estimation of GPU application performance on machines with increased GPU-link performance ratios without knowing the internal structure, enabling effective resource allocation and optimization for improved computational performance.

Problems solved by technology

Currently, performance estimation in such a situation needs information about how GPU computation and CPU-GPU communication depend on each other in the target GPU application, which is not so easy to obtain and determine.

Images