Fix #209: improve socket transport performance
  Split transfers over multiple sockets
  Launch multiple threads to drive sockets
  Detect AWS NICs and set nsockets/nthreads accordingly


[ROCm/rccl commit: 7c72dee660]
This commit is contained in:
Ke Wen
2019-06-25 13:22:47 -07:00
parent 00bc3a724a
commit b91d8170f8
7 changed files with 425 additions and 96 deletions
-6
View File
@@ -13,11 +13,6 @@
extern ncclNet_t* ncclNet;
typedef char ncclNetHandle_t[NCCL_NET_HANDLE_MAXSIZE];
/* Socket Interface Selection type */
typedef enum { findSubnetIf = -1,
dontCareIf = -2
} ncclSocketIfSl_t;
// Translation to external API
static const char* ncclNetName() { return ncclNet->name; }
static ncclResult_t ncclNetDevices(int* ndev) { NCCLCHECK(ncclNet->devices(ndev)); return ncclSuccess; }
@@ -36,7 +31,6 @@ static ncclResult_t ncclNetCloseSend(void* sendComm) { NCCLCHECK(ncclNet->closeS
static ncclResult_t ncclNetCloseRecv(void* recvComm) { NCCLCHECK(ncclNet->closeRecv(recvComm)); return ncclSuccess; }
static ncclResult_t ncclNetCloseListen(void* listenComm) { NCCLCHECK(ncclNet->closeListen(listenComm)); return ncclSuccess; }
extern ncclResult_t ncclSocketCreateHandle(void* opaqueHandle, const char* str);
extern ncclNet_t ncclNetIb;
extern ncclNet_t ncclNetSocket;