Merge pull request #138 from wenkaidu/slice_steps
Revert collective chunk and slice steps to avoid drop in throughput
Этот коммит содержится в:
@@ -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
|
||||
|
||||
Ссылка в новой задаче
Block a user