Optimization for Tree allreduce on A100.
Improve aggregation performance.
Use shared buffers for inter-node send/recv.
Add NVTX profiling hooks.
Accelerate alltoall connections by merging communication for all
channels.
Add support for one hop communication through NVLink, for faster
send/recv communication on cubemesh topologies like DGX-1.
Improve alltoall scheduling to better balance intra/inter node
communication.
Increase send/recv parallelism by 8x, each warp sending or
receiving to a different peer.
Net: move to v4.
Net: make flush operation asynchronous to accelerate alltoall.
Net: define maximum number of requests.
Fix hang when using LL128 protocol after 2^31 steps.
Fix #379 : topology injection failing when using less GPUs than
described in the XML.
Fix #394 : protocol mismatch causing hangs or crashes when using
one GPU per node.
This commit is contained in:
Sylvain Jeaugey
2020-09-04 14:35:05 -07:00
parent 084207e685
commit 920dbe5b35
90 changed files with 11172 additions and 3209 deletions
+16 -3
View File
@@ -48,8 +48,8 @@ struct ncclRecvMem {
struct {
uint64_t tail;
char pad1[CACHE_LINE_SIZE-sizeof(uint64_t)];
char pad2[CACHE_LINE_SIZE-sizeof(uint64_t)];
int sizesFifo[NCCL_STEPS];
void* ptrsFifo[NCCL_STEPS];
};
char pad4[MEM_ALIGN];
};
@@ -63,6 +63,10 @@ struct ncclComm {
struct ncclTopoSystem* topo;
void* bootstrap;
// Bitmasks for ncclTransportP2pSetup
int connect;
uint32_t* connectSend;
uint32_t* connectRecv;
int rank; // my rank in the communicator
int nRanks; // number of GPUs in communicator
@@ -127,7 +131,7 @@ struct ncclComm {
int* intraCudaDevs;
int* intraCGMode; // Whether we can use CUDA9 CGMD or not
int* intraCC; // Only to check all have the same ComputeCap and disable CGMode if not
struct ncclColl args;
struct ncclWorkElem args;
void* argsptr;
// Global proxy thread
@@ -136,8 +140,17 @@ struct ncclComm {
// Whether this communicator uses collNet
int collNetSupport;
// Store info of async operations
struct ncclInfo* asyncOps;
int asyncOpCount;
size_t asyncTotalSize;
//list of async p2p operation queued in a group semantics
struct ncclP2Plist p2plist;
struct ncclP2Plist* p2pSends;
struct ncclP2Plist* p2pRecvs;
int p2pSendCount;
int p2pRecvCount;
};
#endif