【知识】cuda检测GPU是否支持P2P通信及一些注意事项

转载请注明出处：小锋学长生活大爆炸[xfxuezhagn.cn]

如果本文帮助到了你，欢迎[点赞、收藏、关注]哦~

代码流程

1. 先检查所有GPU之间是否支持P2P通信；
2. 然后尝试启用GPU之间的P2P通信；
3. 再次检查所有GPU之间是否支持P2P通信。

test.cu：

#include <stdio.h>
#include <cuda_runtime.h>void checkP2P(int deviceCount) {printf("\nChecking initial P2P support:\n");for (int i = 0; i < deviceCount; i++) {for (int j = 0; j < deviceCount; j++) {if (i != j) {int canAccessPeer = 0;cudaDeviceCanAccessPeer(&canAccessPeer, i, j);printf("GPU %d -> GPU %d: %s\n", i, j, canAccessPeer ? "Supported" : "Not Supported");}}}
}void enableP2P(int deviceCount) {printf("\nEnabling P2P:\n");for (int i = 0; i < deviceCount; i++) {cudaSetDevice(i);for (int j = 0; j < deviceCount; j++) {if (i != j) {int canAccessPeer = 0;cudaDeviceCanAccessPeer(&canAccessPeer, i, j);if (canAccessPeer) {cudaError_t err = cudaDeviceEnablePeerAccess(j, 0);if (err == cudaSuccess) {printf("P2P enabled between GPU %d and GPU %d.\n", i, j);} else {printf("Failed to enable P2P between GPU %d and GPU %d: %s\n", i, j, cudaGetErrorString(err));}}}}}
}void recheckP2P(int deviceCount) {printf("\nRechecking P2P support after enabling:\n");for (int i = 0; i < deviceCount; i++) {for (int j = 0; j < deviceCount; j++) {if (i != j) {int canAccessPeer = 0;cudaDeviceCanAccessPeer(&canAccessPeer, i, j);printf("GPU %d -> GPU %d: %s\n", i, j, canAccessPeer ? "Enabled" : "Disabled");}}}
}int main() {int deviceCount;cudaGetDeviceCount(&deviceCount);if (deviceCount < 2) {printf("At least two GPUs are required for P2P testing.\n");return 0;}printf("Detected %d GPUs.\n", deviceCount);// Step 1: Check initial P2P supportcheckP2P(deviceCount);// Step 2: Enable P2PenableP2P(deviceCount);// Step 3: Recheck P2P supportrecheckP2P(deviceCount);return 0;
}

编译：

nvcc -o test test.cu

执行：

./test.cu

---

注意事项

1、官方提供的cuda_tool里的simpleP2P、p2pBandwidthLatencyTest也都可以检测是否支持P2P。

2、Bridges间不支持P2P通信：deploying-managing-gpu-clusters

3、相同PCIe complex下是支持P2P：ParallelComputing

---

部分不支持P2P的类型

Nvidia Confirms GeForce Cards Lack P2P