NCCL 2.27.3-1
Symmetric memory API and symmetric kernels * Redesign from the ground up, enabling major latency and bandwidth improvements. * Add new API calls to register user-allocated memory among communicator ranks into a NCCL window: ncclCommWindowRegister() and ncclCommWindowDeregister(). The calls currently support symmetric registration for P2P and NVLS, and require VMM memory buffers (i.e., CUMEM must be operational). * Implement specialized kernels taking advantage of symmetrically registered memory, with performance gains expected particularly for small to medium message sizes. * The kernels support 32 bit floating point types and smaller, and sum as the reduction operator, with no more than one collective operation per group. * Floating point summation is always done in fp32 accumulators (with the exception of fp8 on NVLS, where it uses fp16 inside the switch). Thus, the accuracy with fp8 and fp16 data types should be much improved. * This initial implementation supports non-network communicators only (P2P and NVLS transports). * To explore this functionality users need to use the new memory registration API calls with the NCCL_WIN_COLL_SYMMETRIC flag and all ranks of a communicator must pass buffers at the same offset in the same registration when invoking a collective NCCL operation. Add support for DGX Spark. Add support for DirectNIC (CX8) to the internal IB plugin. Add a new ncclCommShrink() API call * It is a non-collective call similar to ncclCommSplit(), which makes it possible to exclude some (possibly unresponsive) ranks from the parent communicator. Add support for loading multiple network plugins * This enables the creation of generic containers that can work across a range of providers. * Allow NCCL_NET_PLUGIN to accept a comma-separated list of plugins to load. NVLink SHARP (NVLS) improvements * Implement NVLS+IB SHARP support for AllGather and ReduceScatter with user buffer registration. This improves performance and reduces the number of CTAs needed to achieve peak bandwidth. * Gracefully fall back by default to other transports if NVLS initialization fails (the old behavior of returning an error code from a NCCL call can be preserved by setting NCCL_NVLS_ENABLE=1). * Decrease the NVLS channel count to 24 on Blackwell systems with multiple NVLink domains per communicator. * Enable fine-tuning of NCCL behavior per communicator using new "ncclConfig_t" members "collnetEnable", "CTAPolicy", and "nvlsCTAs". Profiler improvements * Extend the init function by adding communicator name, comm id (hash), rank, number of ranks, number of nodes, and the NCCL log function to the argument list. This makes the name and the comm id available to all events in the communicator without explicitly passing them to each individual event. Add the communicator id and rank to the profiler trace filename. Now, the communicator name can be set via a new "ncclConfig_t" member "commName". * Improve the accuracy of the GPU kernel events by providing GPU-generated timestamps for the start and stop of every NCCL operation. * Harmonize proxy events, removing overlaps between ProxyOp and ProxyStep states. * Add support for network-defined event updates (through "recordEventState"). * Report the correct number of channels used by every collective/p2p operation (used to be set to nMaxChannels for collectives and absent for p2ps). * Fix the logic on proxyCtrl Idle/Active events (Issue #1162). * Fix an issue where the network proxy profiler could lose track of an event identifier (Issue #1682). * Improve the backward compatibility with plugins older than v4. * Ensure that the work counters are 0-initialized. * Fix a potential race condition in the network profiler that could result in an event being linked to a wrong parent. MNNVL improvements * Increase to 16 the number of NICs used to communicate between MNNVL domains on GB200 systems, to optimize the performance of collective operations. * Add support for more complex MNNVL topologies with up to 32 NICs per node. * If the MNNVL fabric initialization was unsuccessful, NCCL will now fail by default, so as to avoid inadvertently falling back to a potentially much slower network transport. Such failures are typically due to a misconfigured IMEX support on the system. To continue without MNNVL, restart the job with NCCL_MNNVL_ENABLE=0. * Fix a potential hang in alltoall-like communication patterns at a scale of over 80 ranks. * Make NCCL_P2P_DISABLE=1 imply NCCL_MNNVL_ENABLE=0 (so the latter no longer needs to be specified on MNNVL systems). * Fix an initialization failure when NCCL_TOPO_FILE is used on MNNVL systems. * Fix the graph search to exclude non-local NICs. * Fix the SHM transport to use fabric handles on MNNVL systems. NIC Fusion improvements * Disable the creation of fused NICs for physical devices that haven't been merged. * Flatten multiple ports to a single PCI device within the internal IB plugin and reparent dual-port NICs under the first PCI parent. If the parent is not a PCI switch, PCI devices for fused NICs won't be duplicated. * Route traffic on GB200-CX8 systems through DirectNIC, not the host interface. Improve support for platforms with C2C connectivity (e.g., GB200) * Enable GPUDirect RDMA for the NICs by default. * Add support for P2C (PXN over C2C) and the LL128 protocol. Extend NCCL fault tolerance in multithreaded scenarios * Support the creation of multiple nonblocking communicators within a single group and polling in parallel for the completion using multiple threads (one per communicator). Enable ncclImplicitOrderLaunch for CUDA 12.9+ * This can potentially speed up NCCL_IMPLICIT_LAUNCH_ORDER. Improve the netSocket transport latency and control * Provide finer control over the size of the socket send/receive buffers, the task size, and the number of sockets that a single peer can open. * Add support for the inlining of small messages behind the header when using multiple sockets per connection. Improve the readability of the CPU affinity in the debug output * Print it as a range string rather than a bitmask. Fix a potential race condition in graph execution * A contention could arise when mixing graph and non-graph execution. Improve PXN connection code * Avoid duplicate and unused connections. RAS fixes * Fix a memory corruption at job termination time in case of a previously failed initialization of a RAS socket connection. * Fix a race condition leading to a crash when generating a RAS report during communicator initialization (Issues #1669, #1718). * Fix a potential race condition when gathering data for a RAS status report. Fix a potential memory corruption in ncclCommSplit() * Memory could get corrupted when resource sharing was in use and the size of the NVLink domain in the new communicator was smaller than in the old one. Fix asynchronous graph upload * Fix a small memory leak. * Fix oversychronization. Add a check for out-of-memory conditions in ncclMemAlloc() Clean up the NCCL socket code * accept() will retry also if just reading the magic failed (Issue #1613). * connect() will retry also if poll() did not return a POLLOUT event (Issue #1618). * Add error checking in a few instances (Issue #1539). * Fix the loop condition in ncclFindInterfaceMatchSubnet() (Issue #1574). * Clean up the debug output, downgrading WARN messages to INFO in non-critical cases, and printing the peer's address where relevant. Switch NCCL_DEBUG_FILE to line buffering * This should help avoid mixed-up partial output lines in multithreaded cases. Other minor fixes * Improve the checks for buffer overflows in the graph code (Issue #1585). * Extend logging and state clearing to all four events in the internal IB plugin (Issue #1650). * Fix the error path in case IB communication is not ready (Issue #1489). * Add ECE logging for IB fabric. * Fix various minor issues in the graph module (Issue #1635). * Clean up the debug output in the graph code, downgrading WARN messages to INFO in non-critical cases. * Add a missing argument to a directSend() call (Issue #1628). * Remove duplicate code in sendProxySetup() (Issue #1420). * Fix the order of arguments of cudaDeviceCanAccessPeer() (Issue #1507). * Fix compiler warnings with GCC 14. * Fix a typo in a comment (Issue #1236).
Este cometimento está contido em:
Kamil Iskra
+62
-36
@@ -214,7 +214,7 @@ ncclResult_t ncclGetLevel(int* level, const char* disableEnv, const char* levelE
|
||||
const char* str = ncclGetEnv(disableEnv);
|
||||
if (str) {
|
||||
int disable = strtol(str, NULL, 0);
|
||||
if (disable == 1) l = 0;
|
||||
if (disable == 1) l = PATH_LOC;
|
||||
if (l >= 0) INFO(NCCL_ALL, "%s set by environment to %d", disableEnv, disable);
|
||||
}
|
||||
}
|
||||
@@ -247,7 +247,18 @@ ncclResult_t ncclGetLevel(int* level, const char* disableEnv, const char* levelE
|
||||
|
||||
NCCL_PARAM(IgnoreDisabledP2p, "IGNORE_DISABLED_P2P", 0);
|
||||
|
||||
int ncclTopoUserP2pLevel = -1;
|
||||
static int ncclTopoUserP2pLevel = -1; // Initially "uninitialized". When initialized but unset, changes to -2.
|
||||
|
||||
// Gets the user-provided value of NCCL_P2P_LEVEL/NCCL_P2P_DISABLE. If the user did not provide any, the value
|
||||
// of the "level" argument is left unchanged.
|
||||
ncclResult_t ncclGetUserP2pLevel(int* level) {
|
||||
if (ncclTopoUserP2pLevel == -1)
|
||||
NCCLCHECK(ncclGetLevel(&ncclTopoUserP2pLevel, "NCCL_P2P_DISABLE", "NCCL_P2P_LEVEL"));
|
||||
if (ncclTopoUserP2pLevel != -2)
|
||||
*level = ncclTopoUserP2pLevel;
|
||||
return ncclSuccess;
|
||||
}
|
||||
|
||||
ncclResult_t ncclTopoCheckP2p(struct ncclComm* comm, struct ncclTopoSystem* system, int rank1, int rank2,
|
||||
int* p2p, int *read, int* intermediateRank) {
|
||||
int mnnvl = 0;
|
||||
@@ -275,9 +286,9 @@ ncclResult_t ncclTopoCheckP2p(struct ncclComm* comm, struct ncclTopoSystem* syst
|
||||
|
||||
// Get GPUs from topology
|
||||
int g1, g2;
|
||||
NCCLCHECK(ncclTopoRankToIndex(system, rank1, &g1));
|
||||
NCCLCHECK(ncclTopoRankToIndex(system, rank1, &g1, /*showWarn=*/true));
|
||||
struct ncclTopoNode* gpu1 = system->nodes[GPU].nodes+g1;
|
||||
if (ncclTopoRankToIndex(system, rank2, &g2) == ncclInternalError) {
|
||||
if (ncclTopoRankToIndex(system, rank2, &g2, /*showWarn=*/false) == ncclInternalError) {
|
||||
// GPU not found, we can't use p2p.
|
||||
return ncclSuccess;
|
||||
}
|
||||
@@ -302,15 +313,8 @@ ncclResult_t ncclTopoCheckP2p(struct ncclComm* comm, struct ncclTopoSystem* syst
|
||||
if ((arch == NCCL_TOPO_CPU_ARCH_X86 && vendor == NCCL_TOPO_CPU_VENDOR_AMD) && system->nodes[GPU].count <= 2) p2pLevel = PATH_SYS;
|
||||
|
||||
// User override
|
||||
if (ncclTopoUserP2pLevel == -1)
|
||||
NCCLCHECK(ncclGetLevel(&ncclTopoUserP2pLevel, "NCCL_P2P_DISABLE", "NCCL_P2P_LEVEL"));
|
||||
if (ncclTopoUserP2pLevel != -2) {
|
||||
p2pLevel = ncclTopoUserP2pLevel;
|
||||
goto compare;
|
||||
}
|
||||
NCCLCHECK(ncclGetUserP2pLevel(&p2pLevel));
|
||||
|
||||
|
||||
compare:
|
||||
// Compute the PCI distance and compare with the p2pLevel.
|
||||
if (path->type <= p2pLevel) *p2p = 1;
|
||||
|
||||
@@ -378,7 +382,8 @@ NCCL_PARAM(NetGdrRead, "NET_GDR_READ", -2);
|
||||
int ncclTopoUserGdrLevel = -1;
|
||||
const char* ncclTopoGdrModeStr[ncclTopoGdrModeNum] = { "Disabled", "Default", "PCI" };
|
||||
|
||||
NCCL_PARAM(NetGdrC2c, "NET_GDR_C2C", 0);
|
||||
// On C2C platforms use GDRDMA on NICs which are connected to the CPUs
|
||||
NCCL_PARAM(NetGdrC2c, "NET_GDR_C2C", 1);
|
||||
|
||||
ncclResult_t ncclTopoCheckGdr(struct ncclTopoSystem* system, int rank, int64_t netId, int read, enum ncclTopoGdrMode* gdrMode) {
|
||||
*gdrMode = ncclTopoGdrModeDisable;
|
||||
@@ -387,7 +392,7 @@ ncclResult_t ncclTopoCheckGdr(struct ncclTopoSystem* system, int rank, int64_t n
|
||||
int n, g;
|
||||
NCCLCHECK(ncclTopoIdToIndex(system, NET, netId, &n));
|
||||
struct ncclTopoNode* net = system->nodes[NET].nodes+n;
|
||||
NCCLCHECK(ncclTopoRankToIndex(system, rank, &g));
|
||||
NCCLCHECK(ncclTopoRankToIndex(system, rank, &g, /*showWarn=*/true));
|
||||
struct ncclTopoNode* gpu = system->nodes[GPU].nodes+g;
|
||||
|
||||
// Check that both the NIC and GPUs support it
|
||||
@@ -423,29 +428,29 @@ ncclResult_t ncclTopoCheckGdr(struct ncclTopoSystem* system, int rank, int64_t n
|
||||
// In case of PXN, use the intermediate GPU distance instead
|
||||
int proxyRank;
|
||||
NCCLCHECK(ncclTopoGetIntermediateRank(system, gpu->gpu.rank, netId, &proxyRank));
|
||||
NCCLCHECK(ncclTopoRankToIndex(system, proxyRank, &g));
|
||||
NCCLCHECK(ncclTopoRankToIndex(system, proxyRank, &g, /*showWarn=*/true));
|
||||
gpu = system->nodes[GPU].nodes+g;
|
||||
distance = gpu->paths[NET][n].type;
|
||||
}
|
||||
|
||||
int c;
|
||||
NCCLCHECK(ncclGetLocalCpu(system, g, &c));
|
||||
if (ncclParamNetGdrC2c() && distance == PATH_PHB && gpu->paths[CPU][c].type == PATH_C2C) {
|
||||
// On C2C platforms we can still use GDRDMA on NICs connected to the CPUs
|
||||
INFO(NCCL_NET, "GPU %d / HCA %lx connected to CPU %d via C2C link", rank, netId, c);
|
||||
// On C2C platforms we can still use GDRDMA on NICs connected to the CPUs
|
||||
if (ncclParamNetGdrC2c() && distance == PATH_P2C) {
|
||||
INFO(NCCL_GRAPH | NCCL_NET, "GPU %d / HCA %lx connected via C2C link", rank, netId);
|
||||
distance = PATH_C2C;
|
||||
}
|
||||
|
||||
if (distance > netGdrLevel) {
|
||||
INFO(NCCL_NET,"GPU Direct RDMA Disabled for GPU %d / HCA %lx (distance %d > %d)", rank, netId, distance, netGdrLevel);
|
||||
INFO(NCCL_GRAPH|NCCL_NET,"GPU Direct RDMA Disabled for GPU %d / HCA %lx (distance %d > %d)", rank, netId, distance, netGdrLevel);
|
||||
return ncclSuccess;
|
||||
}
|
||||
|
||||
// Force PCIe mapping if path goes through PCI on a C2C system
|
||||
int c;
|
||||
NCCLCHECK(ncclGetLocalCpu(system, g, &c));
|
||||
if (gpu->paths[CPU][c].type == PATH_C2C && distance != PATH_C2C) *gdrMode = ncclTopoGdrModePci;
|
||||
else *gdrMode = ncclTopoGdrModeDefault;
|
||||
|
||||
INFO(NCCL_NET,"GPU Direct RDMA Enabled for GPU %d / HCA %lx (distance %d <= %d), read %d mode %s", rank, netId, distance, netGdrLevel, read, ncclTopoGdrModeStr[*gdrMode]);
|
||||
INFO(NCCL_GRAPH|NCCL_NET,"GPU Direct RDMA Enabled for GPU %d / HCA %lx (distance %d <= %d), read %d mode %s", rank, netId, distance, netGdrLevel, read, ncclTopoGdrModeStr[*gdrMode]);
|
||||
return ncclSuccess;
|
||||
}
|
||||
|
||||
@@ -480,7 +485,7 @@ ncclResult_t ncclTopoNeedFlush(struct ncclComm* comm, int64_t netId, int netDev,
|
||||
if (props.forceFlush == 1 || ncclParamNetForceFlush()) return ncclSuccess;
|
||||
int g;
|
||||
struct ncclTopoSystem* system = comm->topo;
|
||||
NCCLCHECK(ncclTopoRankToIndex(system, rank, &g));
|
||||
NCCLCHECK(ncclTopoRankToIndex(system, rank, &g, /*showWarn=*/true));
|
||||
struct ncclTopoNode* gpu = system->nodes[GPU].nodes+g;
|
||||
// Flush is required on Ampere and earlier
|
||||
if (gpu->gpu.cudaCompCap >= 90) *flush = 0;
|
||||
@@ -506,8 +511,8 @@ ncclResult_t ncclTopoCheckNet(struct ncclTopoSystem* system, int rank1, int rank
|
||||
*net = 1;
|
||||
// First check the current GPU-to-GPU speed.
|
||||
int g1, g2;
|
||||
if (ncclTopoRankToIndex(system, rank1, &g1) != ncclSuccess ||
|
||||
ncclTopoRankToIndex(system, rank2, &g2) != ncclSuccess) {
|
||||
if (ncclTopoRankToIndex(system, rank1, &g1, /*showWarn=*/false) != ncclSuccess ||
|
||||
ncclTopoRankToIndex(system, rank2, &g2, /*showWarn=*/false) != ncclSuccess) {
|
||||
return ncclSuccess;
|
||||
}
|
||||
|
||||
@@ -533,7 +538,7 @@ ncclResult_t ncclTopoGetIntermediateRank(struct ncclTopoSystem* system, int rank
|
||||
// Get GPU and NET
|
||||
int n, g;
|
||||
NCCLCHECK(ncclTopoIdToIndex(system, NET, netId, &n));
|
||||
NCCLCHECK(ncclTopoRankToIndex(system, rank, &g));
|
||||
NCCLCHECK(ncclTopoRankToIndex(system, rank, &g, /*showWarn=*/true));
|
||||
struct ncclTopoNode* gpu = system->nodes[GPU].nodes+g;
|
||||
struct ncclTopoLinkList* path = gpu->paths[NET]+n;
|
||||
if (path->type == PATH_PXN) {
|
||||
@@ -601,6 +606,8 @@ ncclResult_t ncclTopoGetPxnRanks(struct ncclComm* comm, int** intermediateRanks,
|
||||
return ncclSuccess;
|
||||
}
|
||||
|
||||
NCCL_PARAM(PxnC2c, "PXN_C2C", 0);
|
||||
|
||||
ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclComm* comm) {
|
||||
// Precompute paths between GPUs/NICs.
|
||||
|
||||
@@ -659,6 +666,20 @@ ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclComm
|
||||
}
|
||||
}
|
||||
}
|
||||
// update the GPU -> NIC path in the case of C2C + PHB
|
||||
for (int n = 0; n < system->nodes[NET].count; n++) {
|
||||
struct ncclTopoNode* netNode = system->nodes[NET].nodes + n;
|
||||
for (int g = 0; g < system->nodes[GPU].count; g++) {
|
||||
struct ncclTopoNode* gpuNode = system->nodes[GPU].nodes + g;
|
||||
int c;
|
||||
NCCLCHECK(ncclGetLocalCpu(system, g, &c));
|
||||
if (c == -1) continue;
|
||||
if (gpuNode->paths[NET][n].type == PATH_PHB && gpuNode->paths[CPU][c].type == PATH_C2C) {
|
||||
gpuNode->paths[NET][n].type = PATH_P2C;
|
||||
netNode->paths[GPU][g].type = PATH_P2C;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Update paths for NICs (no GPU Direct, PXN, ...)
|
||||
for (int n=0; n<system->nodes[NET].count; n++) {
|
||||
@@ -674,15 +695,20 @@ ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclComm
|
||||
// PXN = PCI + NVLink.
|
||||
struct ncclTopoNode* peerNode = system->nodes[GPU].nodes+localGpuIndex;
|
||||
// Only use PXN for NIC n if remote GPU p ...
|
||||
if (peerNode->paths[NET][n].type <= PATH_PXB && // Is connected to the NIC through PCI
|
||||
peerNode->paths[GPU][g].type <= PATH_NVL && // Is connected to us through NVLink
|
||||
NCCL_TOPO_ID_SYSTEM_ID(peerNode->id) == NCCL_TOPO_ID_SYSTEM_ID(gpu->id) && // Is on the same node as us
|
||||
(peerNode->paths[NET][n].bw > gpu->paths[NET][n].bw || // Has either higher BW to that NIC
|
||||
gpu->paths[NET][n].type > PATH_PXB)) // or avoids going through a CPU
|
||||
// We can use that GPU as relay to communicate with that NIC.
|
||||
// Only enabling it in the GPU->NIC direction for now to favor
|
||||
// receiving locally and sending remotely (consistent with net.cc)
|
||||
NCCLCHECK(addInterStep(system, GPU, localGpuIndex, GPU, g, NET, n));
|
||||
if (/* (1) is either connected to the NIC with PXB*/
|
||||
(peerNode->paths[NET][n].type <= PATH_PXB ||
|
||||
/* or with P2C and PxN over C2C is enabled */
|
||||
(ncclParamPxnC2c() && peerNode->paths[NET][n].type == PATH_P2C)) &&
|
||||
/* and (2) is connected to us through NVLink */
|
||||
peerNode->paths[GPU][g].type <= PATH_NVL &&
|
||||
/* and (3) is on the same node as us */
|
||||
NCCL_TOPO_ID_SYSTEM_ID(peerNode->id) == NCCL_TOPO_ID_SYSTEM_ID(gpu->id) &&
|
||||
/* and (4) has either higher bw to that NIC or avoid going through the CPU*/
|
||||
(peerNode->paths[NET][n].bw > gpu->paths[NET][n].bw || gpu->paths[NET][n].type > PATH_PXB))
|
||||
// We can use that GPU as relay to communicate with that NIC.
|
||||
// Only enabling it in the GPU->NIC direction for now to favor
|
||||
// receiving locally and sending remotely (consistent with net.cc)
|
||||
NCCLCHECK(addInterStep(system, GPU, localGpuIndex, GPU, g, NET, n));
|
||||
}
|
||||
}
|
||||
if (gpu->paths[NET][n].type < PATH_PHB) {
|
||||
@@ -761,7 +787,7 @@ static ncclResult_t ncclTopoGetNchannels(struct ncclComm* comm, int g /*local gp
|
||||
int peer;
|
||||
struct ncclTopoSystem* system = comm->topo;
|
||||
struct ncclTopoLinkList* path = NULL;
|
||||
if (ncclTopoRankToIndex(system, peerRank, &peer) == ncclSuccess) {
|
||||
if (ncclTopoRankToIndex(system, peerRank, &peer, /*showWarn=*/false) == ncclSuccess) {
|
||||
// Same rank
|
||||
if (g == peer) {
|
||||
*nChannels = -1;
|
||||
|
||||
Criar uma nova questão referindo esta
Bloquear um utilizador