The purpose of this micro-benchmark is to measure the PCIe Bandwidth (Download) in a Multi-GPU setup using both pinned and non-pinned memory.
The benchmark allows you to specify the GPUs as well as the CPU cores to use. Both CPUs and GPUs are required to be specified using a comma-separated list. This allows to understand the NUMA effects in a multi-socket system. clinfo and cat /proc/cpuinfo allow you to understand the topology of GPUs.
--cores -c Comma-separated list of CPU cores to use (Default=0)
--gpu_list -g Comma-separated list of GPUs to use (Default=0)
--help -h Print Help Message
--iterations -i Number of Iterations (Default=10)
--pinned -p Use Pinned Memory (Default=1|Yes)
--sizes -s Number of different CL::Buffer sizes doubling from 256MB (Default=2)
Examples:
Use default GPU (1st GPU) and CPU core
./PCIeBandwidth
Use GPUs 0 and 1 and CPU cores 0 and 2
./PCIeBandwidth -g 0,1 -c 0,2
Use GPUs 0-7 and CPU cores 0-3, 8-11
./PCIeBandwidth -g 0,1,2,3,4,5,6,7 -c 0,1,2,3,8,9,11,12