Optimized primitives for collective multi-GPU communication.
NCCL (pronounced "Nickel") is a stand-alone library of standard collective communication routines, such as all-gather, reduce, broadcast, etc., that have been optimized to achieve high bandwidth over PCIe. NCCL supports up to eight GPUs and can be used in either single- or multi-process (e.g., MPI) applications.
At present, the library implements the following collectives:
- all-reduce
- all-gather
- reduce-scatter
- reduce
- broadcast
These collectives are implemented using ring algorithms and have been optimized primarily for throughput. For best performance, small collectives should be batched into larger operations whenever possible. Small test binaries demonstrating how to use each of the above collectives are also provided.
NCCL requires at least CUDA 7.0 and Kepler or newer GPUs. Best performance is achieved when all GPUs are located on a common PCIe root complex, but multi-socket configurations are also supported.
Note: NCCL may also work with CUDA 6.5, but this is an untested configuration.
To build the library and tests.
$ cd nccl
$ make CUDA_HOME=<cuda install path> test
Test binaries are located in the subdirectories nccl/build/test and nccl/build/mpitest.
$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:./build/lib
$ ./build/test/allreduce_test
Error: must specify at least data size in bytes!
Tests nccl AllReduce with user supplied arguments.
Usage: all_reduce_test <data size in bytes> [number of GPUs] [GPU 0] [GPU 1] ...
$ ./build/test/allreduce_test 10000000
# Using devices
# Device 0 -> 0 [0x0a] GeForce GTX TITAN X
# Device 1 -> 1 [0x09] GeForce GTX TITAN X
# Device 2 -> 2 [0x06] GeForce GTX TITAN X
# Device 3 -> 3 [0x05] GeForce GTX TITAN X
# out-of-place in-place
# bytes N type op time algbw busbw res time algbw busbw res
10000000 10000000 char sum 1.628 6.14 9.21 0e+00 1.932 5.18 7.77 0e+00
10000000 10000000 char prod 1.629 6.14 9.21 0e+00 1.643 6.09 9.13 0e+00
10000000 10000000 char max 1.621 6.17 9.25 0e+00 1.634 6.12 9.18 0e+00
10000000 10000000 char min 1.633 6.12 9.19 0e+00 1.637 6.11 9.17 0e+00
10000000 2500000 int sum 1.611 6.21 9.31 0e+00 1.626 6.15 9.23 0e+00
10000000 2500000 int prod 1.613 6.20 9.30 0e+00 1.629 6.14 9.21 0e+00
10000000 2500000 int max 1.619 6.18 9.26 0e+00 1.627 6.15 9.22 0e+00
10000000 2500000 int min 1.619 6.18 9.27 0e+00 1.624 6.16 9.24 0e+00
10000000 5000000 half sum 1.617 6.18 9.28 4e-03 1.636 6.11 9.17 4e-03
10000000 5000000 half prod 1.618 6.18 9.27 1e-03 1.657 6.03 9.05 1e-03
10000000 5000000 half max 1.608 6.22 9.33 0e+00 1.621 6.17 9.25 0e+00
10000000 5000000 half min 1.610 6.21 9.32 0e+00 1.627 6.15 9.22 0e+00
10000000 2500000 float sum 1.618 6.18 9.27 5e-07 1.622 6.17 9.25 5e-07
10000000 2500000 float prod 1.614 6.20 9.29 1e-07 1.628 6.14 9.21 1e-07
10000000 2500000 float max 1.616 6.19 9.28 0e+00 1.633 6.12 9.19 0e+00
10000000 2500000 float min 1.613 6.20 9.30 0e+00 1.628 6.14 9.21 0e+00
10000000 1250000 double sum 1.629 6.14 9.21 0e+00 1.628 6.14 9.21 0e+00
10000000 1250000 double prod 1.619 6.18 9.26 2e-16 1.628 6.14 9.21 2e-16
10000000 1250000 double max 1.613 6.20 9.30 0e+00 1.630 6.13 9.20 0e+00
10000000 1250000 double min 1.622 6.16 9.25 0e+00 1.623 6.16 9.24 0e+00
To install, run make PREFIX=<install dir> install
and add <instal dir>/lib
to your LD_LIBRARY_PATH
.
NCCL follows the MPI collectives API fairly closely. Before any collectives can be called, a communicator object must be initialized on each GPU. On a single-process machine, all GPUs can be conveniently initialized using ncclCommInitAll
. For multi-process applications (e.g., with MPI), ncclCommInitRank
must be called for each GPU. Internally ncclCommInitRank
invokes a synchronization among all GPUs, so these calls must be invoked in different host threads (or processes) for each GPU. A brief single-process example follows, for an MPI example see src/mpi_test.cu. For details about the API see nccl.h.
#include <nccl.h>
typedef struct {
double* sendBuff;
double* recvBuff;
int size;
cudaStream_t stream;
} PerThreadData;
int main(int argc, char* argv[])
{
int nGPUs;
cudaGetDeviceCount(&nGPUs);
ncclComm_t* comms = (ncclComm_t*)malloc(sizeof(ncclComm_t)*nGPUs);
ncclCommInitAll(comms, nGPUs); // initialize communicator
// One communicator per process
PerThreadData* data;
... // Allocate data and issue work to each GPU's
// perDevStream to populate the sendBuffs.
for(int i=0; i<nGPUs; ++i) {
cudaSetDevice(i); // Correct device must be set
// prior to each collective call.
ncclAllReduce(data[i].sendBuff, data[i].recvBuff, size,
ncclDouble, ncclSum, comms[i], data[i].stream);
}
... // Issue work into data[*].stream to consume buffers, etc.
}