Network user buffer support for collectives
* Leverage user buffer registration to achieve zero-copy
inter-node communications for Ring, NVLS and Collnet
Add RAS subsystem
* Create a RAS thread keeping track of all NCCL communicators.
* Add a ncclras tool contacting the RAS thread and getting a
report.
Add fp8 support
* Add support for e5m2 and e4m3 8-bit floating point operations.
* Use Tree/PAT algorithms when possible for better numerical
stability.
Add NIC fusion
* Add a NET API to ask the network plugin to fuse a set of
interfaces together.
* Fuse multiple NICs under the same PCI switch as a single,
larger NIC.
Socket connection failure retry
* Retry in case of socket connection failure (unreachable host)
* Avoid "Software caused connection abort" errors on retries
QP connection failure retry
* Retry in case of IB QP connection failure during ibv_modify_qp.
NET API improvements
* Allow plugins to force a flush in case data and completion
ordering is not guaranteed.
* Indicate when completion is not needed (e.g. for the LL128
protocol), allowing plugins to skip generating a completion.
* Allow for full offload of allgather operations when using one
GPU per node.
NCCL_ALGO/NCCL_PROTO strict enforcement
* Extend NCCL_ALGO/NCCL_PROTO syntax to be able to specify
ALGO/PROTO filters for each collective operation.
* Strictly enforce the ALGO/PROTO filters, no longer fall back
on the ring algorithm when the filtering leaves no option and
error out instead.
Enable CUMEM host allocations
* Use cumem functions for host memory allocation by default.
Improved profiler plugin API
* Avoid dependencies with NCCL includes.
* Add information on whether the buffer is registered or not
Adjust PAT tuning
* Improve transition between PAT and ring at scale.
Fix hangs when running with different CPU architectures
* Detect when we use a mix of GPU architectures
* Ensure Algo/Proto decisions are made based on that unified
state.
Fix FD leak in UDS
* Fix a leak when mapping buffers intra-node with cumem IPCs.
Fix crash when mixing buffer registration and graph buffer registration.
* Separate local and graph registration to avoid crashes when we free
buffers.
Fix user buffer registration with dmabuf
* Make ncclSend/ncclRecv communication with buffer registration functional
on network plugins relying on dmabuf for buffer registration.
Fix crash in IB code caused by uninitialized fields.
Fix non-blocking ncclSend/ncclRecv
* Fix case where ncclSend/ncclRecv would return ncclSuccess in non-blocking
mode even though the operation was not enqueued onto the stream.
* Issue #1495
Various compiler tweaks and fixes
* PR #758
Fix typo in ncclTopoPrintGraph
* Issue #1468
[ROCm/rccl commit: 6aae379278]
232 satır
6.4 KiB
C++
232 satır
6.4 KiB
C++
/*
|
|
* Copyright (c) 2016-2023, NVIDIA CORPORATION. All rights reserved.
|
|
*
|
|
* See COPYRIGHT for license information
|
|
*/
|
|
|
|
#include "ipcsocket.h"
|
|
#include "utils.h"
|
|
#include <stdlib.h>
|
|
#include <string.h>
|
|
#include <errno.h>
|
|
|
|
// Enable Linux abstract socket naming
|
|
#define USE_ABSTRACT_SOCKET
|
|
|
|
#define NCCL_IPC_SOCKNAME_STR "/tmp/nccl-socket-%d-%lx"
|
|
|
|
/*
|
|
* Create a Unix Domain Socket
|
|
*/
|
|
ncclResult_t ncclIpcSocketInit(ncclIpcSocket *handle, int rank, uint64_t hash, volatile uint32_t* abortFlag) {
|
|
int fd = -1;
|
|
struct sockaddr_un cliaddr;
|
|
char temp[NCCL_IPC_SOCKNAME_LEN] = "";
|
|
|
|
if (handle == NULL) {
|
|
return ncclInternalError;
|
|
}
|
|
|
|
handle->fd = -1;
|
|
handle->socketName[0] = '\0';
|
|
if ((fd = socket(AF_UNIX, SOCK_DGRAM, 0)) < 0) {
|
|
WARN("UDS: Socket creation error : %s (%d)", strerror(errno), errno);
|
|
return ncclSystemError;
|
|
}
|
|
|
|
bzero(&cliaddr, sizeof(cliaddr));
|
|
cliaddr.sun_family = AF_UNIX;
|
|
|
|
// Create unique name for the socket.
|
|
int len = snprintf(temp, NCCL_IPC_SOCKNAME_LEN, NCCL_IPC_SOCKNAME_STR, rank, hash);
|
|
if (len > (sizeof(cliaddr.sun_path) - 1)) {
|
|
WARN("UDS: Cannot bind provided name to socket. Name too large");
|
|
close(fd);
|
|
return ncclInternalError;
|
|
}
|
|
#ifndef USE_ABSTRACT_SOCKET
|
|
unlink(temp);
|
|
#endif
|
|
|
|
TRACE(NCCL_INIT, "UDS: Creating socket %s", temp);
|
|
|
|
strncpy(cliaddr.sun_path, temp, len);
|
|
#ifdef USE_ABSTRACT_SOCKET
|
|
cliaddr.sun_path[0] = '\0'; // Linux abstract socket trick
|
|
#endif
|
|
if (bind(fd, (struct sockaddr *)&cliaddr, sizeof(cliaddr)) < 0) {
|
|
WARN("UDS: Binding to socket %s failed : %s (%d)", temp, strerror(errno), errno);
|
|
close(fd);
|
|
return ncclSystemError;
|
|
}
|
|
|
|
handle->fd = fd;
|
|
strcpy(handle->socketName, temp);
|
|
|
|
handle->abortFlag = abortFlag;
|
|
// Mark socket as non-blocking
|
|
if (handle->abortFlag) {
|
|
int flags;
|
|
SYSCHECK(flags = fcntl(fd, F_GETFL), "fcntl");
|
|
SYSCHECK(fcntl(fd, F_SETFL, flags | O_NONBLOCK), "fcntl");
|
|
}
|
|
|
|
return ncclSuccess;
|
|
}
|
|
|
|
ncclResult_t ncclIpcSocketGetFd(struct ncclIpcSocket* handle, int* fd) {
|
|
if (handle == NULL) {
|
|
WARN("ncclSocketGetFd: pass NULL socket");
|
|
return ncclInvalidArgument;
|
|
}
|
|
if (fd) *fd = handle->fd;
|
|
return ncclSuccess;
|
|
}
|
|
|
|
ncclResult_t ncclIpcSocketClose(ncclIpcSocket *handle) {
|
|
if (handle == NULL) {
|
|
return ncclInternalError;
|
|
}
|
|
if (handle->fd <= 0) {
|
|
return ncclSuccess;
|
|
}
|
|
#ifndef USE_ABSTRACT_SOCKET
|
|
if (handle->socketName[0] != '\0') {
|
|
unlink(handle->socketName);
|
|
}
|
|
#endif
|
|
close(handle->fd);
|
|
|
|
return ncclSuccess;
|
|
}
|
|
|
|
ncclResult_t ncclIpcSocketRecvMsg(ncclIpcSocket *handle, void *hdr, int hdrLen, int *recvFd) {
|
|
struct msghdr msg = {0, 0, 0, 0, 0, 0, 0};
|
|
struct iovec iov[1];
|
|
|
|
// Union to guarantee alignment requirements for control array
|
|
union {
|
|
struct cmsghdr cm;
|
|
char control[CMSG_SPACE(sizeof(int))];
|
|
} control_un;
|
|
|
|
struct cmsghdr *cmptr;
|
|
char dummy_buffer[1];
|
|
int ret;
|
|
|
|
msg.msg_control = control_un.control;
|
|
msg.msg_controllen = sizeof(control_un.control);
|
|
|
|
if (hdr == NULL) {
|
|
iov[0].iov_base = (void *)dummy_buffer;
|
|
iov[0].iov_len = sizeof(dummy_buffer);
|
|
} else {
|
|
iov[0].iov_base = hdr;
|
|
iov[0].iov_len = hdrLen;
|
|
}
|
|
|
|
msg.msg_iov = iov;
|
|
msg.msg_iovlen = 1;
|
|
|
|
while ((ret = recvmsg(handle->fd, &msg, 0)) <= 0) {
|
|
if (errno != EAGAIN && errno != EWOULDBLOCK && errno != EINTR) {
|
|
WARN("UDS: Receiving data over socket failed : %d", errno);
|
|
return ncclSystemError;
|
|
}
|
|
if (handle->abortFlag && __atomic_load_n(handle->abortFlag, __ATOMIC_ACQUIRE)) return ncclInternalError;
|
|
}
|
|
|
|
if (recvFd != NULL) {
|
|
if (((cmptr = CMSG_FIRSTHDR(&msg)) != NULL) && (cmptr->cmsg_len == CMSG_LEN(sizeof(int)))) {
|
|
if ((cmptr->cmsg_level != SOL_SOCKET) || (cmptr->cmsg_type != SCM_RIGHTS)) {
|
|
WARN("UDS: Receiving data over socket failed");
|
|
return ncclSystemError;
|
|
}
|
|
|
|
memmove(recvFd, CMSG_DATA(cmptr), sizeof(*recvFd));
|
|
} else {
|
|
WARN("UDS: Receiving data over socket %s failed", handle->socketName);
|
|
return ncclSystemError;
|
|
}
|
|
TRACE(NCCL_INIT|NCCL_P2P, "UDS: Got recvFd %d from socket %s", *recvFd, handle->socketName);
|
|
}
|
|
|
|
return ncclSuccess;
|
|
}
|
|
|
|
ncclResult_t ncclIpcSocketRecvFd(ncclIpcSocket *handle, int *recvFd) {
|
|
return ncclIpcSocketRecvMsg(handle, NULL, 0, recvFd);
|
|
}
|
|
|
|
ncclResult_t ncclIpcSocketSendMsg(ncclIpcSocket *handle, void *hdr, int hdrLen, const int sendFd, int rank, uint64_t hash) {
|
|
struct msghdr msg = {0, 0, 0, 0, 0, 0, 0};
|
|
struct iovec iov[1];
|
|
char temp[NCCL_IPC_SOCKNAME_LEN];
|
|
|
|
union {
|
|
struct cmsghdr cm;
|
|
char control[CMSG_SPACE(sizeof(int))];
|
|
} control_un;
|
|
|
|
struct cmsghdr *cmptr;
|
|
char dummy_buffer[1];
|
|
struct sockaddr_un cliaddr;
|
|
|
|
// Construct client address to send this shareable handle to
|
|
bzero(&cliaddr, sizeof(cliaddr));
|
|
cliaddr.sun_family = AF_UNIX;
|
|
|
|
int len = snprintf(temp, NCCL_IPC_SOCKNAME_LEN, NCCL_IPC_SOCKNAME_STR, rank, hash);
|
|
if (len > (sizeof(cliaddr.sun_path) - 1)) {
|
|
WARN("UDS: Cannot connect to provided name for socket. Name too large");
|
|
return ncclInternalError;
|
|
}
|
|
(void) strncpy(cliaddr.sun_path, temp, len);
|
|
|
|
#ifdef USE_ABSTRACT_SOCKET
|
|
cliaddr.sun_path[0] = '\0'; // Linux abstract socket trick
|
|
#endif
|
|
|
|
TRACE(NCCL_INIT, "UDS: Sending hdr %p len %d fd %d to UDS socket %s", hdr, hdrLen, sendFd, temp);
|
|
|
|
if (sendFd != -1) {
|
|
msg.msg_control = control_un.control;
|
|
msg.msg_controllen = sizeof(control_un.control);
|
|
|
|
cmptr = CMSG_FIRSTHDR(&msg);
|
|
cmptr->cmsg_len = CMSG_LEN(sizeof(int));
|
|
cmptr->cmsg_level = SOL_SOCKET;
|
|
cmptr->cmsg_type = SCM_RIGHTS;
|
|
memmove(CMSG_DATA(cmptr), &sendFd, sizeof(sendFd));
|
|
}
|
|
|
|
msg.msg_name = (void *)&cliaddr;
|
|
msg.msg_namelen = sizeof(struct sockaddr_un);
|
|
|
|
if (hdr == NULL) {
|
|
iov[0].iov_base = (void *)dummy_buffer;
|
|
iov[0].iov_len = sizeof(dummy_buffer);
|
|
} else {
|
|
iov[0].iov_base = hdr;
|
|
iov[0].iov_len = hdrLen;
|
|
}
|
|
msg.msg_iov = iov;
|
|
msg.msg_iovlen = 1;
|
|
msg.msg_flags = 0;
|
|
|
|
ssize_t sendResult;
|
|
while ((sendResult = sendmsg(handle->fd, &msg, 0)) < 0) {
|
|
if (errno != EAGAIN && errno != EWOULDBLOCK && errno != EINTR) {
|
|
WARN("UDS: Sending data over socket %s failed : %s (%d)", temp, strerror(errno), errno);
|
|
return ncclSystemError;
|
|
}
|
|
if (handle->abortFlag && __atomic_load_n(handle->abortFlag, __ATOMIC_ACQUIRE)) return ncclInternalError;
|
|
}
|
|
|
|
return ncclSuccess;
|
|
}
|
|
|
|
ncclResult_t ncclIpcSocketSendFd(ncclIpcSocket *handle, const int sendFd, int rank, uint64_t hash) {
|
|
return ncclIpcSocketSendMsg(handle, NULL, 0, sendFd, rank, hash);
|
|
}
|