Merge pull request #138 from wenkaidu/slice_steps

Revert collective chunk and slice steps to avoid drop in throughput
Этот коммит содержится в:
Wenkai Du
2019-10-18 13:30:27 -07:00
коммит произвёл GitHub
родитель a95529a6e2 df74d12946
Коммит 998ab83675
+12 -6
Просмотреть файл
@@ -57,12 +57,18 @@
DECL_ALL_COLLS
// CHUNKSIZE must be a multiple of SLICESIZE
#define ALLREDUCE_SLICESTEPS (NCCL_STEPS/4)
#define ALLREDUCE_CHUNKSTEPS (NCCL_STEPS/2)
#define ALLGATHER_SLICESTEPS (NCCL_STEPS/4)
#define ALLGATHER_CHUNKSTEPS (NCCL_STEPS/2)
#define REDUCESCATTER_SLICESTEPS (NCCL_STEPS/4)
#define REDUCESCATTER_CHUNKSTEPS (NCCL_STEPS/2)
//#define ALLREDUCE_SLICESTEPS (NCCL_STEPS/4)
//#define ALLREDUCE_CHUNKSTEPS (NCCL_STEPS/2)
//#define ALLGATHER_SLICESTEPS (NCCL_STEPS/4)
//#define ALLGATHER_CHUNKSTEPS (NCCL_STEPS/2)
//#define REDUCESCATTER_SLICESTEPS (NCCL_STEPS/4)
//#define REDUCESCATTER_CHUNKSTEPS (NCCL_STEPS/2)
#define ALLREDUCE_SLICESTEPS 4
#define ALLREDUCE_CHUNKSTEPS 4
#define ALLGATHER_SLICESTEPS 4
#define ALLGATHER_CHUNKSTEPS 4
#define REDUCESCATTER_SLICESTEPS 4
#define REDUCESCATTER_CHUNKSTEPS 4
#define BROADCAST_SLICESTEPS 1
#define BROADCAST_CHUNKSTEPS 1
#define REDUCE_SLICESTEPS 1