作者｜KIDGINBROOK更新｜潘麗晨上次介紹到rank0的機(jī)器生成了ncclUniqueId，并完成了機(jī)器的bootstrap網(wǎng)絡(luò)和通信網(wǎng)絡(luò)的初始化，這節(jié)接著看下所有節(jié)點(diǎn)間bootstrap的連接是如何建立的。

rank0節(jié)點(diǎn)執(zhí)行ncclGetUniqueId生成ncclUniqueId，通過(guò)mpi將Id廣播到所有節(jié)點(diǎn)，然后所有節(jié)點(diǎn)都會(huì)執(zhí)行ncclCommInitRank，這里其他節(jié)點(diǎn)也會(huì)進(jìn)行初始化bootstrap網(wǎng)絡(luò)和通信網(wǎng)絡(luò)的操作，然后會(huì)執(zhí)行到ncclCommInitRankSync。

(資料圖片僅供參考)

ncclResult_t ncclCommInitRankSync(ncclComm_t* newcomm, int nranks, ncclUniqueId commId, int myrank, int cudaDev) { ncclResult_t res; CUDACHECK(cudaSetDevice(cudaDev)); NCCLCHECKGOTO(commAlloc(newcomm, nranks, myrank), res, cleanup); NCCLCHECKGOTO(initTransportsRank(*newcomm, &commId), res, cleanup); NCCLCHECKGOTO(devCommSetup(*newcomm), res, cleanup); INFO(NCCL_INIT,"comm %p rank %d nranks %d cudaDev %d busId %x - Init COMPLETE", *newcomm, myrank, nranks, (*newcomm)->cudaDev, (*newcomm)->busId); return ncclSuccess;cleanup: if ((*newcomm) && (*newcomm)->bootstrap) bootstrapAbort((*newcomm)->bootstrap); *newcomm = NULL; return res;}

ncclComm_t是指向ncclComm的指針，ncclComm是一個(gè)大雜燴，包含了通信用到的所有上下文信息，里面的字段等用到的時(shí)候再介紹，然后通過(guò)commAlloc分配newcom，并且完成初始化，比如當(dāng)前是哪個(gè)卡，對(duì)應(yīng)的pcie busid是什么，然后執(zhí)行initTransportsRank。

static ncclResult_t initTransportsRank(struct ncclComm* comm, ncclUniqueId* commId) { // We use 3 AllGathers // 1. { peerInfo, comm } // 2. ConnectTransport[nranks], ConnectValue[nranks] // 3. { nThreads, nrings, compCap, prev[MAXCHANNELS], next[MAXCHANNELS] } int rank = comm->rank; int nranks = comm->nRanks; uint64_t commHash = getHash(commId->internal, NCCL_UNIQUE_ID_BYTES); TRACE(NCCL_INIT, "comm %p, commHash %lx, rank %d nranks %d - BEGIN", comm, commHash, rank, nranks); NCCLCHECK(bootstrapInit(commId, rank, nranks, &comm->bootstrap)); // AllGather1 - begin struct { struct ncclPeerInfo peerInfo; struct ncclComm* comm; } *allGather1Data; NCCLCHECK(ncclCalloc(&allGather1Data, nranks)); allGather1Data[rank].comm = comm; struct ncclPeerInfo* myInfo = &allGather1Data[rank].peerInfo; NCCLCHECK(fillInfo(comm, myInfo, commHash)); NCCLCHECK(bootstrapAllGather(comm->bootstrap, allGather1Data, sizeof(*allGather1Data))); NCCLCHECK(ncclCalloc(&comm->peerInfo, nranks+1)); // Extra rank to represent CollNet root for (int i = 0; i < nranks; i++) { memcpy(comm->peerInfo+i, &allGather1Data[i].peerInfo, sizeof(struct ncclPeerInfo)); if ((i != rank) && (comm->peerInfo[i].hostHash == myInfo->hostHash) && (comm->peerInfo[i].busId == myInfo->busId)) { WARN("Duplicate GPU detected : rank %d and rank %d both on CUDA device %x", rank, i, myInfo->busId); return ncclInvalidUsage; } }

看下bootstrapInit：

ncclResult_t bootstrapInit(ncclUniqueId * id, int rank, int nranks, void** commState) { ncclNetHandle_t* netHandle = (ncclNetHandle_t*) id; bool idFromEnv = getenv("NCCL_COMM_ID") != NULL; struct extState* state; NCCLCHECK(ncclCalloc(&state, 1)); state->rank = rank; state->nranks = nranks; *commState = state; TRACE(NCCL_INIT, "rank %d nranks %d", rank, nranks); struct extInfo info = { 0 }; info.rank = rank; info.nranks = nranks; void *tmpSendComm, *tmpRecvComm; // Pass the remote address to listen via info if (idFromEnv) { memcpy(&info.extHandleListen, netHandle, sizeof(ncclNetHandle_t)); memcpy(&info.extHandleListenRoot, netHandle, sizeof(ncclNetHandle_t)); } // listen will return the local address via info (specify interface type "findSubnetIf") state->dev = idFromEnv ? findSubnetIf : 0; void* extBstrapListenCommRoot; NCCLCHECK(bootstrapNetListen(state->dev, &info.extHandleListen, &state->extBstrapListenComm)); NCCLCHECK(bootstrapNetListen(state->dev, &info.extHandleListenRoot, &extBstrapListenCommRoot)); // stagger connection times to avoid an overload of the root at very high rank counts if (nranks > 128) { long msec = rank; struct timespec tv; tv.tv_sec = msec / 1000; tv.tv_nsec = 1000000 * (msec % 1000); TRACE(NCCL_INIT, "rank %d delaying connection to root by %ld msec", rank, msec); (void) nanosleep(&tv, NULL); } // send info on my listening socket to root NCCLCHECK(bootstrapNetConnect(state->dev, netHandle, &tmpSendComm)); NCCLCHECK(bootstrapNetSend(tmpSendComm, &info, sizeof(info))); NCCLCHECK(bootstrapNetCloseSend(tmpSendComm)); // get info on my "next" rank in the bootstrap ring from root}

首先看下commState，即ncclComm的bootstrap，類型為extState。

struct extState { void* extBstrapListenComm; void* extBstrapRingRecvComm; void* extBstrapRingSendComm; ncclNetHandle_t* peerBstrapHandles; struct unexConn* unexpectedConnections; int rank; int nranks; int dev;};

其中extBstrapRingSendComm是當(dāng)前節(jié)點(diǎn)連接next的socket連接，extBstrapRingRecvComm是當(dāng)前節(jié)點(diǎn)和prev節(jié)點(diǎn)的socket連接，extBstrapListenComm是當(dāng)前節(jié)點(diǎn)的監(jiān)聽socket，peerBstrapHandles是所有rank的ip port（對(duì)應(yīng)extBstrapListenComm），dev默認(rèn)為0，表示用第幾個(gè)ip地址。

然后通過(guò)bootstrapNetListen創(chuàng)建extHandleListen和extHandleListenRoot兩個(gè)bootstrap comm，如前文所述，bootstrap comm其實(shí)就是保存了fd，這里創(chuàng)建兩個(gè)comm的原因是extHandleListen是rank之間實(shí)際使用的bootstrap連接，extHandleListenRoot是rank0節(jié)點(diǎn)和其他所有rank進(jìn)行通信使用的連接。

static ncclResult_t bootstrapNetListen(int dev, ncclNetHandle_t* netHandle, void** listenComm)

bootstrapNetListen函數(shù)上節(jié)有介紹過(guò)，會(huì)獲取到第dev個(gè)當(dāng)前機(jī)器的ip，然后listen獲取監(jiān)聽fd，將ip port寫到nethandle，獲取到的bootstrap comm寫到listencomm。

然后將rank，nrank，extHandleListen和extHandleListenRoot寫到extInfo里。

struct extInfo { int rank; int nranks; ncclNetHandle_t extHandleListenRoot; ncclNetHandle_t extHandleListen;};

netHandle為ncclUniqueId，即rank0的ip port，然后通過(guò)bootstrapNetConnect創(chuàng)建bootstrap send comm，類比bootstrapNetListen，bootstrapNetConnect就是建立到netHandle的socket連接，將socket寫到sendComm里，這里dev并沒(méi)有用到。

static?ncclResult_t?bootstrapNetConnect(int?dev,?ncclNetHandle_t*?netHandle,?void**?sendComm)

然后通過(guò)bootstrapNetSend將extInfo發(fā)送出去，即發(fā)給rank0：

static ncclResult_t bootstrapNetSend(void* sendComm, void* data, int size) { struct bootstrapNetComm* comm = (struct bootstrapNetComm*)sendComm; NCCLCHECK(socketSend(comm->fd, &size, sizeof(int))); NCCLCHECK(socketSend(comm->fd, data, size)); return ncclSuccess;}

其中socketSend就是執(zhí)行send接口發(fā)送數(shù)據(jù)。

然后通過(guò)bootstrapNetCloseSend關(guān)閉fd。

rank0收到數(shù)據(jù)后會(huì)做什么工作呢，回顧一下，rank0的節(jié)執(zhí)行ncclGetUniqueId生成ncclUniqueId，其中在執(zhí)行bootstrapCreateRoot的最后會(huì)啟動(dòng)一個(gè)線程執(zhí)行bootstrapRoot。

static void *bootstrapRoot(void* listenComm) { struct extInfo info; ncclNetHandle_t *rankHandles = NULL; ncclNetHandle_t *rankHandlesRoot = NULL; // for initial rank <-> root information exchange ncclNetHandle_t zero = { 0 }; // for sanity checking void* tmpComm; ncclResult_t res; setFilesLimit(); TRACE(NCCL_INIT, "BEGIN"); /* Receive addresses from all ranks */ int nranks = 0, c = 0; do { NCCLCHECKGOTO(bootstrapNetAccept(listenComm, &tmpComm), res, out); NCCLCHECKGOTO(bootstrapNetRecv(tmpComm, &info, sizeof(info)), res, out); NCCLCHECKGOTO(bootstrapNetCloseRecv(tmpComm), res, out); if (c == 0) { nranks = info.nranks; NCCLCHECKGOTO(ncclCalloc(&rankHandles, nranks), res, out); NCCLCHECKGOTO(ncclCalloc(&rankHandlesRoot, nranks), res, out); } if (nranks != info.nranks) { WARN("Bootstrap Root : mismatch in rank count from procs %d : %d", nranks, info.nranks); goto out; } if (memcmp(&zero, &rankHandlesRoot[info.rank], sizeof(ncclNetHandle_t)) != 0) { WARN("Bootstrap Root : rank %d of %d ranks has already checked in", info.rank, nranks); goto out; } // Save the connection handle for that rank memcpy(rankHandlesRoot+info.rank, info.extHandleListenRoot, sizeof(ncclNetHandle_t)); memcpy(rankHandles+info.rank, info.extHandleListen, sizeof(ncclNetHandle_t)); ++c; TRACE(NCCL_INIT, "Received connect from rank %d total %d/%d", info.rank, c, nranks); } while (c < nranks); TRACE(NCCL_INIT, "COLLECTED ALL %d HANDLES", nranks); // Send the connect handle for the next rank in the AllGather ring for (int r=0; r

listenComm是上一個(gè)博文中rank0創(chuàng)建的監(jiān)聽fd，bootstrapNetAccept是從listenComm中獲取一個(gè)新連接，使用新連接的fd創(chuàng)建recvcomm。

static ncclResult_t bootstrapNetAccept(void* listenComm, void** recvComm)

然后通過(guò)bootstrapNetRecv讀取tmpComm的數(shù)據(jù)，即其他rank發(fā)送來(lái)的extInfo，然后保存其他rank的extHandleListen和extHandleListenRoot，這個(gè)時(shí)候rank0就獲取到其他所有rank的ip和port了。 獲取完所有rank的info之后開始建環(huán)，將節(jié)點(diǎn)(r+1) % nranks的extHandleListen發(fā)送給節(jié)點(diǎn)r，就是說(shuō)將節(jié)點(diǎn)r的next節(jié)點(diǎn)的nethandle發(fā)送給節(jié)點(diǎn)r。這里可以看出，每個(gè)節(jié)點(diǎn)創(chuàng)建了兩個(gè)listen comm，其中rank0使用extHandleListenRoot進(jìn)行通信，其他節(jié)點(diǎn)之間通過(guò)extHandleListen進(jìn)行通信。

然后再回去接著看bootstrapInit。

ncclResult_t bootstrapInit(ncclUniqueId * id, int rank, int nranks, void** commState) { // get info on my "next" rank in the bootstrap ring from root ncclNetHandle_t extHandleNext; NCCLCHECK(bootstrapNetAccept(extBstrapListenCommRoot, &tmpRecvComm)); NCCLCHECK(bootstrapNetRecv(tmpRecvComm, &extHandleNext, sizeof(extHandleNext))); NCCLCHECK(bootstrapNetCloseRecv(tmpRecvComm)); NCCLCHECK(bootstrapNetCloseListen(extBstrapListenCommRoot)); NCCLCHECK(bootstrapNetConnect(state->dev, &extHandleNext, &state->extBstrapRingSendComm)); // Accept the connect request from the previous rank in the AllGather ring NCCLCHECK(bootstrapNetAccept(state->extBstrapListenComm, &state->extBstrapRingRecvComm)); // AllGather all listen handlers NCCLCHECK(ncclCalloc(&state->peerBstrapHandles, nranks)); memcpy(state->peerBstrapHandles+rank, info.extHandleListen, sizeof(ncclNetHandle_t)); NCCLCHECK(bootstrapAllGather(state, state->peerBstrapHandles, sizeof(ncclNetHandle_t))); TRACE(NCCL_INIT, "rank %d nranks %d - DONE", rank, nranks); return ncclSuccess;}

接著所有rank都會(huì)在extHandleListenRoot上接收新連接創(chuàng)建tmpRecvComm，然后接收到當(dāng)前rank的next的ip，port；然后連接next創(chuàng)建bscomm到state->extBstrapRingSendComm，接收prev的連接創(chuàng)建bscomm到state->extBstrapRingRecvComm，到現(xiàn)在bootstrap網(wǎng)絡(luò)連接就完全建立起來(lái)了，如下圖：

最后gather所有rank的ip port，首先將自己的nethandle放到peerBstrapHandles的對(duì)應(yīng)位置，如下所示。

然后執(zhí)行bootstrapAllGather：

ncclResult_t bootstrapAllGather(void* commState, void* allData, int size) { struct extState* state = (struct extState*)commState; char* data = (char*)allData; int rank = state->rank; int nranks = state->nranks; TRACE(NCCL_INIT, "rank %d nranks %d size %d", rank, nranks, size); /* Simple ring based AllGather * At each step i receive data from (rank-i-1) from left * and send previous step"s data from (rank-i) to right */ for (int i=0; iextBstrapRingSendComm, data+sslice*size, size)); // Recv slice from the left NCCLCHECK(bootstrapNetRecv(state->extBstrapRingRecvComm, data+rslice*size, size)); } TRACE(NCCL_INIT, "rank %d nranks %d size %d - DONE", rank, nranks, size); return ncclSuccess;}

每一次將自己的data發(fā)送給對(duì)應(yīng)的rank，然后接收其他rank發(fā)送過(guò)來(lái)的data，如下圖。

第一步：

第二步：

到這里每個(gè)rank就都有了全局所有rank的ip port。

最后總結(jié)一下，本節(jié)主要?jiǎng)?chuàng)建了bootstrap環(huán)形網(wǎng)絡(luò)連接，并保存到ncclComm里。

其他人都在看

“ChatGPT們”的淘金時(shí)代

大型語(yǔ)言模型的推理演算

GPT-4創(chuàng)造者：第二次改變AI浪潮的方向

ChatGPT作者Schulman：我們成功的秘密武器

比快更快，開源Stable Diffusion刷新作圖速度

OneEmbedding:單卡訓(xùn)練TB級(jí)推薦模型不是夢(mèng)

GLM訓(xùn)練加速：性能最高提升3倍，顯存節(jié)省1/3

歡迎Star、試用OneFlow: github.com/Oneflow-Inc/oneflow/http://github.com/Oneflow-Inc/oneflow/