-
Created by
Marco Verlato, last updated on Sept 27, 2022
2 minute read
If you have two or more GPUs, some application can enhance their performance by exploiting P2P high speed connection among the GPUs. The tests below can be executed on bare metal as well as on VMs.
- clone cuda-sample repository and build simpleP2P and p2pBandwidthLatencyTest tests:
[root@cld-dfa-gpu-04]# git clone https://github.com/NVIDIA/cuda-samples.git [root@cld-dfa-gpu-04]# yum -y group install "Development Tools" [root@cld-dfa-gpu-04]# cd cuda-samples/Samples/0_Introduction/simpleP2P; make; cd [root@cld-dfa-gpu-04]# cd cuda-samples/Samples/5_Domain_Specific/p2pBandwidthLatencyTest; make; cd
- execute simpleP2P program:
[root@cld-dfa-gpu-04]# cuda-samples/Samples/0_Introduction/simpleP2P/simpleP2P [cuda-samples/Samples/0_Introduction/simpleP2P/simpleP2P] - Starting... Checking for multiple GPUs... CUDA-capable device count: 2 Checking GPU(s) for support of peer to peer memory access... > Peer access from NVIDIA A2 (GPU0) -> NVIDIA A2 (GPU1) : Yes > Peer access from NVIDIA A2 (GPU1) -> NVIDIA A2 (GPU0) : Yes Enabling peer access between GPU0 and GPU1... Allocating buffers (64MB on GPU0, GPU1 and CPU Host)... Creating event handles... cudaMemcpyPeer / cudaMemcpy between GPU0 and GPU1: 10.56GB/s Preparing host buffer and memcpy to GPU0... Run kernel on GPU1, taking source data from GPU0 and writing to GPU1... Run kernel on GPU0, taking source data from GPU1 and writing to GPU0... Copy data back to host from GPU0 and verify results... Disabling peer access... Shutting down... Test passed
- execute p2pBandwidthLatencyTest program:
[root@cld-dfa-gpu-04]# cuda-samples/Samples/5_Domain_Specific/p2pBandwidthLatencyTest/p2pBandwidthLatencyTest
[P2P (Peer-to-Peer) GPU Bandwidth Latency Test]
Device: 0, NVIDIA A2, pciBusID: 17, pciDeviceID: 0, pciDomainID:0
Device: 1, NVIDIA A2, pciBusID: ca, pciDeviceID: 0, pciDomainID:0
Device=0 CAN Access Peer Device=1
Device=1 CAN Access Peer Device=0
***NOTE: In case a device doesn't have P2P access to other one, it falls back to normal memcopy procedure.
So you can see lesser Bandwidth (GB/s) and unstable Latency (us) in those cases.
P2P Connectivity Matrix
D\D 0 1
0 1 1
1 1 1
Unidirectional P2P=Disabled Bandwidth Matrix (GB/s)
D\D 0 1
0 149.61 11.70
1 11.66 164.30
Unidirectional P2P=Enabled Bandwidth (P2P Writes) Matrix (GB/s)
D\D 0 1
0 149.58 11.36
1 11.36 164.34
Bidirectional P2P=Disabled Bandwidth Matrix (GB/s)
D\D 0 1
0 156.74 16.32
1 16.31 164.48
Bidirectional P2P=Enabled Bandwidth Matrix (GB/s)
D\D 0 1
0 156.74 20.86
1 20.86 164.48
P2P=Disabled Latency Matrix (us)
GPU 0 1
0 1.44 10.70
1 12.97 1.34
CPU 0 1
0 2.72 6.56
1 6.55 2.60
P2P=Enabled Latency (P2P Writes) Matrix (us)
GPU 0 1
0 1.43 1.20
1 1.22 1.33
CPU 0 1
0 2.76 2.01
1 1.96 2.64
NOTE: The CUDA Samples are not meant for performance measurements. Results may vary when GPU Boost is enabled.