Gloo is a collective communications library. It comes with a number of collective algorithms useful for machine learning applications. These include a barrier, broadcast, and allreduce.
Transport of data between participating machines is abstracted so that IP can be used at all times, or InifiniBand (or RoCE) when available. In the latter case, if the InfiniBand transport is used, GPUDirect can be used to accelerate cross machine GPU-to-GPU memory transfers.
Where applicable, algorithms have an implementation that works with system memory buffers, and one that works with NVIDIA GPU memory buffers. In the latter case, it is not necessary to copy memory between host and device; this is taken care of by the algorithm implementations.
Gloo is built to run on Linux and has no hard dependencies other than libstdc++. That said, it will generally only be useful when used in combination with a few optional dependencies below.
Optional dependencies are:
- CUDA and NCCL -- for CUDA aware algorithms, tests, and benchmark
- Google Test -- to build and run tests
- Hiredis -- for coordinating machine rendezvous through Redis
- MPI -- for coordinating machine rendezvous through MPI
Please refer to docs/ for detailed documentation.
You can build Gloo using CMake.
Since it is a library, it is most convenient to vendor it in your own project and include the project root in your own CMake configuration.
Building the tests requires Google Test version 1.8 or higher. On
Ubuntu, this version ships with version 17.10 and up. If you run an
older version, you'll have to install Google Test yourself, and set
the GTEST_ROOT CMake variable.
To build the tests, run:
mkdir -p build
cd build
cmake ../ -DBUILD_TEST=1 -DGTEST_ROOT=/some/path (if using custom install)
make
ls -l gloo/test/gloo_test*To test the CUDA algorithms, specify USE_CUDA=ON as well, and the
CUDA tests are built at gloo/test/gloo_test_cuda.
First install the dependencies required by the benchmark tool. On Ubuntu, you can do so by running:
sudo apt-get install -y libhiredis-devThen build the benchmark, run:
mkdir build
cd build
cmake ../ -DBUILD_BENCHMARK=1
ls -l gloo/benchmark/benchmarkThe benchmark tool depends on Redis/Hiredis for rendezvous. The benchmark tool for CUDA algorithms obviously also depends on both CUDA and NCCL.
To run a benchmark:
-
Copy the benchmark tool to all participating machines
-
Start a Redis server on any host (either a client machine or one of the machines participating in the test). Note that Redis Cluster is not supported.
-
Determine some unique ID for the benchmark run (e.g. the
uuidtool or some number). -
On each machine, run (or pass
--helpfor more options):./benchmark \ --size <number of machines> \ --rank <index of this machine, starting at 0> \ --redis-host <Redis host> \ --redis-port <Redis port> \ --prefix <unique identifier for this run> \ --transport tcp \ --elements <number of elements; -1 for a sweep> \ --iteration-time 1s \ allreduce_ring_chunked
Example output (running on 4 machines with a 40GbE network):
elements min (us) p50 (us) p99 (us) max (us) samples
1 195 263 342 437 3921
2 195 261 346 462 4039
5 197 261 339 402 3963
10 197 263 338 398 3749
20 199 268 343 395 4146
50 200 265 344 401 3889
100 205 265 351 414 3645
200 197 264 328 387 3960
500 201 264 329 394 4274
1000 200 267 330 380 3344
2000 205 263 323 395 3682
5000 240 335 424 460 3277
10000 271 346 402 457 2721
20000 283 358 392 428 2719
50000 342 438 495 649 1654
100000 413 487 669 799 1687
200000 1113 1450 1837 2801 669
500000 1099 1294 1665 1959 560
1000000 1858 2286 2779 6100 320
2000000 3546 3993 4364 4886 252
5000000 10030 10608 11106 11628 92
Gloo is BSD-licensed.