Communicator Benchmark for Chainer

Collection of microbenchmarks for Chainer. Originally developed by Takuya Akiba and maintained by developers.

• comm_bench1: Microbenchmark to test Chainer's communicators' allreduce_grad performance stats, with ResNet gradients
• comm_bench2: Microbenchmark to test Chainer's communicators' allreduce_grad performance in timeseries plot, with ResNet gradients
• comm_bench3: Microbenchmark to test raw latency of allreduce of MPI and NCCL

Usage example & Prerequisites

Requires Python >= 3.5 and MPI which is supported by Chainer.

Install

$ git clone git://github.com/chainer/comm_bench.git
...
$ cd comm_bench
$ pip install --user .

Usage example

$ mpirun --np 8 -- comm_bench1 --communicator_names pure_nccl,hierarchial
...
$ mpirun --np 16 -hostfile hosts.txt -- comm_bench2 --interval 0.2
...

License

MIT License, (C) 2019 Preferred Networks, Inc.