NCCL源码解析①:初始化及ncclUniqueId的产生
作者|KIDGINBROOK
更新|潘丽晨
int e = cmd; \
if( e != MPI_SUCCESS ) { \
printf("Failed: MPI error %s:%d '%d'\n", \
__FILE__,__LINE__, e); \
exit(EXIT_FAILURE); \
} \
} while(0)
cudaError_t e = cmd; \
if( e != cudaSuccess ) { \
printf("Failed: Cuda error %s:%d '%s'\n", \
__FILE__,__LINE__,cudaGetErrorString(e)); \
exit(EXIT_FAILURE); \
} \
} while(0)
ncclResult_t r = cmd; \
if (r!= ncclSuccess) { \
printf("Failed, NCCL error %s:%d '%s'\n", \
__FILE__,__LINE__,ncclGetErrorString(r)); \
exit(EXIT_FAILURE); \
} \
} while(0)
static uint64_t getHostHash(const char* string) {
// Based on DJB2a, result = result * 33 ^ char
uint64_t result = 5381;
for (int c = 0; string[c] != '\0'; c++){
result = ((result << 5) + result) ^ string[c];
}
return result;
}
static void getHostName(char* hostname, int maxlen) {
gethostname(hostname, maxlen);
for (int i=0; i< maxlen; i++) {
if (hostname[i] == '.') {
hostname[i] = '\0';
return;
}
}
}
int main(int argc, char* argv[])
{
int size = 32*1024*1024;
int myRank, nRanks, localRank = 0;
//initializing MPI
MPICHECK(MPI_Init(&argc, &argv));
MPICHECK(MPI_Comm_rank(MPI_COMM_WORLD, &myRank));
MPICHECK(MPI_Comm_size(MPI_COMM_WORLD, &nRanks));
//calculating localRank which is used in selecting a GPU
uint64_t hostHashs[nRanks];
char hostname[1024];
getHostName(hostname, 1024);
hostHashs[myRank] = getHostHash(hostname);
MPICHECK(MPI_Allgather(MPI_IN_PLACE, 0, MPI_DATATYPE_NULL, hostHashs, sizeof(uint64_t), MPI_BYTE, MPI_COMM_WORLD));
for (int p=0; p<nRanks; p++) {
if (p == myRank) break;
if (hostHashs[p] == hostHashs[myRank]) localRank++;
}
//each process is using two GPUs
int nDev = 2;
float** sendbuff = (float**)malloc(nDev * sizeof(float*));
float** recvbuff = (float**)malloc(nDev * sizeof(float*));
cudaStream_t* s = (cudaStream_t*)malloc(sizeof(cudaStream_t)*nDev);
//picking GPUs based on localRank
for (int i = 0; i < nDev; ++i) {
CUDACHECK(cudaSetDevice(localRank*nDev + i));
CUDACHECK(cudaMalloc(sendbuff + i, size * sizeof(float)));
CUDACHECK(cudaMalloc(recvbuff + i, size * sizeof(float)));
CUDACHECK(cudaMemset(sendbuff[i], 1, size * sizeof(float)));
CUDACHECK(cudaMemset(recvbuff[i], 0, size * sizeof(float)));
CUDACHECK(cudaStreamCreate(s+i));
}
ncclUniqueId id;
ncclComm_t comms[nDev];
//generating NCCL unique ID at one process and broadcasting it to all
if (myRank == 0) ncclGetUniqueId(&id);
MPICHECK(MPI_Bcast((void *)&id, sizeof(id), MPI_BYTE, 0, MPI_COMM_WORLD));
//initializing NCCL, group API is required around ncclCommInitRank as it is
//called across multiple GPUs in each thread/process
NCCLCHECK(ncclGroupStart());
for (int i=0; i<nDev; i++) {
CUDACHECK(cudaSetDevice(localRank*nDev + i));
NCCLCHECK(ncclCommInitRank(comms+i, nRanks*nDev, id, myRank*nDev + i));
}
NCCLCHECK(ncclGroupEnd());
//calling NCCL communication API. Group API is required when using
//multiple devices per thread/process
NCCLCHECK(ncclGroupStart());
for (int i=0; i<nDev; i++)
NCCLCHECK(ncclAllReduce((const void*)sendbuff[i], (void*)recvbuff[i], size, ncclFloat, ncclSum,
comms[i], s[i]));
NCCLCHECK(ncclGroupEnd());
//synchronizing on CUDA stream to complete NCCL communication
for (int i=0; i<nDev; i++)
CUDACHECK(cudaStreamSynchronize(s[i]));
//freeing device memory
for (int i=0; i<nDev; i++) {
CUDACHECK(cudaFree(sendbuff[i]));
CUDACHECK(cudaFree(recvbuff[i]));
}
//finalizing NCCL
for (int i=0; i<nDev; i++) {
ncclCommDestroy(comms[i]);
}
//finalizing MPI
MPICHECK(MPI_Finalize());
printf("[MPI Rank %d] Success \n", myRank);
return 0;
}
ncclResult_t ncclGetUniqueId(ncclUniqueId* out) {
NCCLCHECK(ncclInit());
NCCLCHECK(PtrCheck(out, "GetUniqueId", "out"));
return bootstrapGetUniqueId(out);
}
ncclResult_t initNet() {
// Always initialize bootstrap network
NCCLCHECK(bootstrapNetInit());
NCCLCHECK(initNetPlugin(&ncclNet, &ncclCollNet));
if (ncclNet != NULL) return ncclSuccess;
if (initNet(&ncclNetIb) == ncclSuccess) {
ncclNet = &ncclNetIb;
} else {
NCCLCHECK(initNet(&ncclNetSocket));
ncclNet = &ncclNetSocket;
}
return ncclSuccess;
}
static int findInterfaces(const char* prefixList, char* names, union socketAddress *addrs, int sock_family, int maxIfNameSize, int maxIfs) {
struct netIf userIfs[MAX_IFS];
bool searchNot = prefixList && prefixList[0] == '^';
if (searchNot) prefixList++;
bool searchExact = prefixList && prefixList[0] == '=';
if (searchExact) prefixList++;
int nUserIfs = parseStringList(prefixList, userIfs, MAX_IFS);
int found = 0;
struct ifaddrs *interfaces, *interface;
getifaddrs(&interfaces);
for (interface = interfaces; interface && found < maxIfs; interface = interface->ifa_next) {
if (interface->ifa_addr == NULL) continue;
int family = interface->ifa_addr->sa_family;
if (family != AF_INET && family != AF_INET6)
continue;
if (sock_family != -1 && family != sock_family)
continue;
if (family == AF_INET6) {
struct sockaddr_in6* sa = (struct sockaddr_in6*)(interface->ifa_addr);
if (IN6_IS_ADDR_LOOPBACK(&sa->sin6_addr)) continue;
}
if (!(matchIfList(interface->ifa_name, -1, userIfs, nUserIfs, searchExact) ^ searchNot)) {
continue;
}
bool duplicate = false;
for (int i = 0; i < found; i++) {
if (strcmp(interface->ifa_name, names+i*maxIfNameSize) == 0) { duplicate = true; break; }
}
if (!duplicate) {
strncpy(names+found*maxIfNameSize, interface->ifa_name, maxIfNameSize);
int salen = (family == AF_INET) ? sizeof(sockaddr_in) : sizeof(sockaddr_in6);
memcpy(addrs+found, interface->ifa_addr, salen);
found++;
}
}
freeifaddrs(interfaces);
return found;
}
ncclResult_t ncclIbInit(ncclDebugLogger_t logFunction) {
static int shownIbHcaEnv = 0;
if(wrap_ibv_symbols() != ncclSuccess) { return ncclInternalError; }
if (ncclParamIbDisable()) return ncclInternalError;
if (ncclNIbDevs == -1) {
pthread_mutex_lock(&ncclIbLock);
wrap_ibv_fork_init();
if (ncclNIbDevs == -1) {
ncclNIbDevs = 0;
if (findInterfaces(ncclIbIfName, &ncclIbIfAddr, MAX_IF_NAME_SIZE, 1) != 1) {
WARN("NET/IB : No IP interface found.");
return ncclInternalError;
}
// Detect IB cards
int nIbDevs;
struct ibv_device** devices;
// Check if user defined which IB device:port to use
char* userIbEnv = getenv("NCCL_IB_HCA");
if (userIbEnv != NULL && shownIbHcaEnv++ == 0) INFO(NCCL_NET|NCCL_ENV, "NCCL_IB_HCA set to %s", userIbEnv);
struct netIf userIfs[MAX_IB_DEVS];
bool searchNot = userIbEnv && userIbEnv[0] == '^';
if (searchNot) userIbEnv++;
bool searchExact = userIbEnv && userIbEnv[0] == '=';
if (searchExact) userIbEnv++;
int nUserIfs = parseStringList(userIbEnv, userIfs, MAX_IB_DEVS);
if (ncclSuccess != wrap_ibv_get_device_list(&devices, &nIbDevs)) return ncclInternalError;
for (int d=0; d<nIbDevs && ncclNIbDevs<MAX_IB_DEVS; d++) {
struct ibv_context * context;
if (ncclSuccess != wrap_ibv_open_device(&context, devices[d]) || context == NULL) {
WARN("NET/IB : Unable to open device %s", devices[d]->name);
continue;
}
int nPorts = 0;
struct ibv_device_attr devAttr;
memset(&devAttr, 0, sizeof(devAttr));
if (ncclSuccess != wrap_ibv_query_device(context, &devAttr)) {
WARN("NET/IB : Unable to query device %s", devices[d]->name);
if (ncclSuccess != wrap_ibv_close_device(context)) { return ncclInternalError; }
continue;
}
for (int port = 1; port <= devAttr.phys_port_cnt; port++) {
struct ibv_port_attr portAttr;
if (ncclSuccess != wrap_ibv_query_port(context, port, &portAttr)) {
WARN("NET/IB : Unable to query port %d", port);
continue;
}
if (portAttr.state != IBV_PORT_ACTIVE) continue;
if (portAttr.link_layer != IBV_LINK_LAYER_INFINIBAND
&& portAttr.link_layer != IBV_LINK_LAYER_ETHERNET) continue;
// check against user specified HCAs/ports
if (! (matchIfList(devices[d]->name, port, userIfs, nUserIfs, searchExact) ^ searchNot)) {
continue;
}
TRACE(NCCL_INIT|NCCL_NET,"NET/IB: [%d] %s:%d/%s ", d, devices[d]->name, port,
portAttr.link_layer == IBV_LINK_LAYER_INFINIBAND ? "IB" : "RoCE");
ncclIbDevs[ncclNIbDevs].device = d;
ncclIbDevs[ncclNIbDevs].guid = devAttr.sys_image_guid;
ncclIbDevs[ncclNIbDevs].port = port;
ncclIbDevs[ncclNIbDevs].link = portAttr.link_layer;
ncclIbDevs[ncclNIbDevs].speed = ncclIbSpeed(portAttr.active_speed) * ncclIbWidth(portAttr.active_width);
ncclIbDevs[ncclNIbDevs].context = context;
strncpy(ncclIbDevs[ncclNIbDevs].devName, devices[d]->name, MAXNAMESIZE);
NCCLCHECK(ncclIbGetPciPath(ncclIbDevs[ncclNIbDevs].devName, &ncclIbDevs[ncclNIbDevs].pciPath, &ncclIbDevs[ncclNIbDevs].realPort));
ncclIbDevs[ncclNIbDevs].maxQp = devAttr.max_qp;
ncclNIbDevs++;
nPorts++;
pthread_create(&ncclIbAsyncThread, NULL, ncclIbAsyncThreadMain, context);
}
if (nPorts == 0 && ncclSuccess != wrap_ibv_close_device(context)) { return ncclInternalError; }
}
if (nIbDevs && (ncclSuccess != wrap_ibv_free_device_list(devices))) { return ncclInternalError; };
}
if (ncclNIbDevs == 0) {
INFO(NCCL_INIT|NCCL_NET, "NET/IB : No device found.");
} else {
char line[1024];
line[0] = '\0';
for (int d=0; d<ncclNIbDevs; d++) {
snprintf(line+strlen(line), 1023-strlen(line), " [%d]%s:%d/%s", d, ncclIbDevs[d].devName,
ncclIbDevs[d].port, ncclIbDevs[d].link == IBV_LINK_LAYER_INFINIBAND ? "IB" : "RoCE");
}
line[1023] = '\0';
char addrline[1024];
INFO(NCCL_INIT|NCCL_NET, "NET/IB : Using%s ; OOB %s:%s", line, ncclIbIfName, socketToString(&ncclIbIfAddr.sa, addrline));
}
pthread_mutex_unlock(&ncclIbLock);
}
return ncclSuccess;
}
ncclResult_t bootstrapCreateRoot(ncclUniqueId* id, bool idFromEnv) {
ncclNetHandle_t* netHandle = (ncclNetHandle_t*) id;
void* listenComm;
NCCLCHECK(bootstrapNetListen(idFromEnv ? dontCareIf : 0, netHandle, &listenComm));
pthread_t thread;
pthread_create(&thread, NULL, bootstrapRoot, listenComm);
return ncclSuccess;
}
static ncclResult_t bootstrapNetListen(int dev, ncclNetHandle_t* netHandle, void** listenComm) {
union socketAddress* connectAddr = (union socketAddress*) netHandle;
static_assert(sizeof(union socketAddress) < NCCL_NET_HANDLE_MAXSIZE, "union socketAddress size is too large");
// if dev >= 0, listen based on dev
if (dev >= 0) {
NCCLCHECK(bootstrapNetGetSocketAddr(dev, connectAddr));
} else if (dev == findSubnetIf) {
...
} // Otherwise, handle stores a local address
struct bootstrapNetComm* comm;
NCCLCHECK(bootstrapNetNewComm(&comm));
NCCLCHECK(createListenSocket(&comm->fd, connectAddr));
*listenComm = comm;
return ncclSuccess;
}
static ncclResult_t bootstrapNetGetSocketAddr(int dev, union socketAddress* addr) {
if (dev >= bootstrapNetIfs) return ncclInternalError;
memcpy(addr, bootstrapNetIfAddrs+dev, sizeof(*addr));
return ncclSuccess;
}
struct bootstrapNetComm {
int fd;
};
static ncclResult_t createListenSocket(int *fd, union socketAddress *localAddr) {
/* IPv4/IPv6 support */
int family = localAddr->sa.sa_family;
int salen = (family == AF_INET) ? sizeof(sockaddr_in) : sizeof(sockaddr_in6);
/* Create socket and bind it to a port */
int sockfd = socket(family, SOCK_STREAM, 0);
if (sockfd == -1) {
WARN("Net : Socket creation failed : %s", strerror(errno));
return ncclSystemError;
}
if (socketToPort(&localAddr->sa)) {
// Port is forced by env. Make sure we get the port.
int opt = 1;
SYSCHECK(setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR | SO_REUSEPORT, &opt, sizeof(opt)), "setsockopt");
SYSCHECK(setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)), "setsockopt");
}
// localAddr port should be 0 (Any port)
SYSCHECK(bind(sockfd, &localAddr->sa, salen), "bind");
/* Get the assigned Port */
socklen_t size = salen;
SYSCHECK(getsockname(sockfd, &localAddr->sa, &size), "getsockname");
char line[1024];
TRACE(NCCL_INIT|NCCL_NET,"Listening on socket %s", socketToString(&localAddr->sa, line));
/* Put the socket in listen mode
* NB: The backlog will be silently truncated to the value in /proc/sys/net/core/somaxconn
*/
SYSCHECK(listen(sockfd, 16384), "listen");
*fd = sockfd;
return ncclSuccess;
}
(本文经授权后由OneFlow发布。原文:https://blog.csdn.net/KIDGIN7439/article/details/126712106?spm=1001.2014.3001.5502)
其他人都在看 点击“阅读原文”,欢迎Star、试用OneFlow新版本
-
InfiniBand组网、RoCE组网、GPU池化管理、算力调度与英伟达SuperPOD的融合发展
InfiniBand组网、RoCE组网、GPU池化管理和算力调度的技术进步,与英伟达SuperPOD的硬件和软件解决方案相结合,共同构建了一个高效、灵活且可扩展的计算环境,为数字化转型和智能化升级提供了坚实的基础。随着技术的不断进步和市场需求的增长,这一领域将继续展现出巨大的潜力和价值。
넶0 2024-11-22 -
VR大空间、大空间LBE、大空间多人互动技术的发展与PICO、HTC VIVE的融合
VR大空间技术、大空间LBE以及大空间多人互动技术的发展,与PICO和HTC VIVE的创新实践相结合,共同推动了VR技术在沉浸式体验领域的进步。随着技术的不断成熟和市场的扩大,这些技术有望为用户带来更加丰富和真实的虚拟体验。
넶0 2024-11-22 -
算力租赁、算力服务、算力中心与英伟达SuperPOD的融合发展
算力租赁、算力服务、算力中心与英伟达SuperPOD的融合发展,共同开启了一个全新的算力时代。这种融合发展模式在各个行业都展现出了巨大的潜力,为各行业的数字化转型和创新发展注入了源源不断的动力,创造了更多的价值和可能。
넶0 2024-11-22 -
InfiniBand组网、RoCE组网、GPU池化管理、算力调度与英伟达SuperPOD的深度融合
InfiniBand组网、RoCE组网、GPU池化管理和算力调度的技术进步,与英伟达SuperPOD的硬件和软件解决方案相结合,共同构建了一个高效、灵活且可扩展的计算环境,为数字化转型和智能化升级提供了坚实的基础。随着技术的不断进步和市场需求的增长,这一领域将继续展现出巨大的潜力和价值。
넶3 2024-11-21 -
VR大空间、大空间LBE、大空间多人互动技术的发展与PICO、HTC VIVE的融合
VR大空间技术、大空间LBE以及大空间多人互动技术的发展,与PICO和HTC VIVE的创新实践相结合,共同推动了VR技术在沉浸式体验领域的进步。随着技术的不断成熟和市场的扩大,这些技术有望为用户带来更加丰富和真实的虚拟体验。
넶3 2024-11-21 -
InfiniBand组网、RoCE组网、GPU池化管理、算力调度与英伟达的协同发展
InfiniBand组网、RoCE组网、GPU池化管理和算力调度的技术进步,与英伟达的SuperPOD解决方案相结合,共同构建了一个高效、灵活且可扩展的计算环境,为数字化转型和智能化升级提供了坚实的基础。随着技术的不断进步和市场需求的增长,这一领域将继续展现出巨大的潜力和价值。
넶4 2024-11-20