NVIDIA NCCL 源码学习（十一）- ring allreduce

之前的章节里我们看到了nccl send/recv通信的过程，本节我们以ring allreduce为例看下集合通信的过程。整体执行流程和send/recv很像，所以对于相似的流程只做简单介绍，主要介绍ring allreduce自己特有内容。

单机

搜索ring

在nccl初始化的过程中会分析机内拓扑，建立CPU，GPU，网卡等PCI节点的拓扑图，并基于这个图搜索一系列的channel，假设单机内执行ncclTopoCompute搜索出的ring为：

graph->intra: GPU/0 GPU/1 GPU/2 GPU/3 GPU/4 GPU/5 GPU/6 GPU/7

接着设置每个channel中ncclRing的prev和next，表示当前rank的前后的GPU，例如GPU0的prev为GPU7，next为GPU1

建链

然后开始建立当前rank到prev和next GPU的链接

for (int c=0; c<comm->nChannels; c++) {
    struct ncclChannel* channel = comm->channels+c;
    NCCLCHECKGOTO(setupChannel(comm, c, rank, nranks, rings+c*nranks), ret, affinity_restore);
    if (comm->nRanks == 1) continue;
    NCCLCHECKGOTO(ncclTransportP2pSetup(comm, &ringGraph, channel, 1, &channel->ring.prev, 1, &channel->ring.next), ret, affinity_restore);
  }

链接建立完成之后如下图所示，buff位于发送端，head和tail为send端recv端共同持有。

为了后续表述方便，这里做一下约定，假设send端为rank0，recv端为rank1，sendbuff指的是用户执行api传进来的输入，recvbuff指的是用户执行api传进来的输出；buffer指的是图一所示的buff；实际rank0将数据发送给rank1的流程是将数据从sendbuff拷贝到位于rank0的buff，recv端接收数据是从位于rank0的buff拷贝到recvbuff，但我们后续会将send流程表述为将sendbuff发送到rank1的buffer，接收流程表述为将数据从当前rank的buffer拷贝到recvbuff（换句话说假设buff位于rank1以方便理解）

图 1

执行api

在完成上述的初始化之后用户开始执行allreduce。

ncclResult_t ncclAllReduce(const void* sendbuff, void* recvbuff, size_t count,
    ncclDataType_t datatype, ncclRedOp_t op, ncclComm* comm, cudaStream_t stream) {
  struct ncclInfo info = { ncclCollAllReduce, "AllReduce",
    sendbuff, recvbuff, count, datatype, op, 0, comm, stream, /* Args */
    ALLREDUCE_CHUNKSTEPS, ALLREDUCE_SLICESTEPS };
  return ncclEnqueueCheck(&info);
}

enqueue

创建info之后执行ncclEnqueueCheck，假设为非group操作

  
  ncclResult_t ncclEnqueueCheck(struct ncclInfo* info) {
  // Launch asynchronously if needed
  if (ncclAsyncMode()) {
    ...
  } else {
    NCCLCHECK(PtrCheck(info->comm, info->opName, "comm"));
    NCCLCHECK(ArgsCheck(info));
    NCCLCHECK(checkSetStream(info));

    INFO(NCCL_COLL,"%s: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p",
        info->opName, info->comm->opCount, info->sendbuff, info->recvbuff, info->count,
        info->datatype, info->op, info->root, info->comm, info->comm->nRanks, info->stream);

    NCCLCHECK(ncclSaveKernel(info));
    NCCLCHECK(ncclBarrierEnqueue(info->comm));
    NCCLCHECK(ncclBarrierEnqueueWait(info->comm));
    NCCLCHECK(ncclEnqueueEvents(info->comm));
    return ncclSuccess;
  }
}

ncclSaveKernel

然后通过ncclSaveKernel将参数等信息添加到channel里。

ncclResult_t ncclSaveKernel(struct ncclInfo* info) {
  ...
  struct ncclColl coll;
  struct ncclProxyArgs proxyArgs;
  memset(&proxyArgs, 0, sizeof(struct ncclProxyArgs));
  NCCLCHECK(computeColl(info, &coll, &proxyArgs));

  info->comm->myParams->blockDim.x = std::max<unsigned>(info->comm->myParams->blockDim.x, info->nThreads);

  int nChannels = info->coll == ncclCollSendRecv ? 1 : coll.args.coll.nChannels;
  int nSubChannels = (info->pattern == ncclPatternCollTreeUp || info->pattern == ncclPatternCollTreeDown) ? 2 : 1;

  for (int bid=0; bid<nChannels*nSubChannels; bid++) {
    int channelId = (info->coll == ncclCollSendRecv) ? info->channelId :
      info->comm->myParams->gridDim.x % info->comm->nChannels;
    struct ncclChannel* channel = info->comm->channels+channelId;

    if (channel->collCount == NCCL_MAX_OPS) {
      WARN("Too many aggregated operations on channel %d (%d max)", channel->id, NCCL_MAX_OPS);
      return ncclInvalidUsage;
    }

    // Proxy
    proxyArgs.channel = channel;
    // Adjust pattern for CollNet based on channel index
    if (nSubChannels == 2) {
      info->pattern = (channelId < info->comm->nChannels/nSubChannels) ? ncclPatternCollTreeUp : ncclPatternCollTreeDown;
    }

    if (info->coll == ncclCollSendRecv) {
      info->comm->myParams->gridDim.x = std::max<unsigned>(info->comm->myParams->gridDim.x, channelId+1);
      NCCLCHECK(ncclProxySaveP2p(info, channel));
    } else {
      NCCLCHECK(ncclProxySaveColl(&proxyArgs, info->pattern, info->root, info->comm->nRanks));
    }
    info->comm->myParams->gridDim.x++;
    int opIndex = channel->collFifoTail;
    struct ncclColl* c = channel->collectives+opIndex;
    volatile uint8_t* activePtr = (volatile uint8_t*)&c->active;
    while (activePtr[0] != 0) sched_yield();

    memcpy(c, &coll, sizeof(struct ncclColl));
    if (info->coll != ncclCollSendRecv) c->args.coll.bid = bid % coll.args.coll.nChannels;

    c->active = 1;
    opIndex = (opIndex+1)%NCCL_MAX_OPS;
    c->nextIndex = opIndex;
    channel->collFifoTail = opIndex;
    channel->collCount++;
  }
  info->comm->opCount++;
  return ncclSuccess;
}

这里核心是computeColl，通过computeColl将kernel所需的参数信息添加到channel的collectives中，并更新myParams->gridDim.x，即一个channel对应一个block。

static ncclResult_t computeColl(struct ncclInfo* info /* input */, struct ncclColl* coll, struct ncclProxyArgs* proxyArgs /* output */) {
  coll->args.sendbuff = info->sendbuff;
  coll->args.recvbuff = info->recvbuff;
  coll->args.comm = info->comm->devComm;

  if (info->coll == ncclCollSendRecv) {
    coll->args.p2p.sendCount = info->sendbytes;
    coll->args.p2p.recvCount = info->recvbytes;
    coll->args.p2p.delta = info->delta;
    coll->funcIndex = FUNC_INDEX_P2P;
    coll->args.p2p.nThreads = info->nThreads = info->comm->maxThreads[NCCL_ALGO_RING][NCCL_PROTO_SIMPLE]+2*WARP_SIZE;
    return ncclSuccess;
  }
  // Set nstepsPerLoop and nchunksPerLoop
  NCCLCHECK(getAlgoInfo(info));
  NCCLCHECK(getPatternInfo(info));
  NCCLCHECK(getLoopInfo(info));
  ...
}  

nccl支持NCCL_PROTO_LL，NCCL_PROTO_LL128和NCCL_PROTO_SIMPLE三种协议，支持NCCL_ALGO_TREE，NCCL_ALGO_RING和NCCL_ALGO_COLLNET三种算法，getAlgoInfo会遍历三种算法和三种协议的组合，选出最好的算法和协议，具体如何选择后边介绍tree allreduce的时候再看，本节先忽略这块逻辑，先假定选出的协议为NCCL_PROTO_SIMPLE，算法为NCCL_ALGO_RING。

getPatternInfo中会将info->pattern设置为ncclPatternRingTwice。

getLoopInfo会设置nstepsPerLoop和nchunksPerLoop。

info->nstepsPerLoop = 2*(info->comm->nRanks-1); info->nchunksPerLoop = info->comm->nRanks;

然后看下上边提到的这几个变量的含义：

图 2

n个rank的ring allreduce过程如上图（图片来自这里）所示，假设要allreduce100M的数据，执行流程会分为多次循环，每次循环执行比如4M的数据，图中展示了allreduce中一次循环的过程，其中(a0+a1+a2+a3)长度为4M，这一次循环过程中分为左侧的reduce scatter和右侧的allgather过程，每个rank每个step收发这块数据的1/n，即1M长度，比如a0，所以nchunksPerLoop = nRanks，表示一次循环过程被分成了多少个数据块，一个数据块叫一个chunk；因为执行了reduce scatter和allgather两次ring的过程，因此pattern叫ncclPatternRingTwice，nstepsPerLoop = 2*(info->comm->nRanks-1)，表示一次循环中需要执行多少步，reduce scatter执行nRank - 1步，allgather也执行nRanks - 1步。

launch kernel

然后执行ncclBarrierEnqueue，核心就是通过setupLaunch将第一个channel的第一个ncclColl拷贝到comm->args，然后通过ncclLaunchCooperativeKernelMultiDevice launch kernel。

ncclResult_t ncclBarrierEnqueue(struct ncclComm* comm) {
  struct cudaLaunchParams* params = comm->myParams;
  if (params->gridDim.x == 0) return ncclSuccess;

  NCCLCHECK(setupLaunch(comm, params));

  // Use internal NCCL stream for CGMD/GROUP launch if required or if the user stream is NULL
  ...
  if (comm->launchMode == ncclComm::GROUP) {
    int isLast = 0;
    NCCLCHECK(ncclCpuBarrierIn(comm, &isLast));
    if (isLast) {
      // I'm the last. Launch all operations.
      NCCLCHECK(ncclLaunchCooperativeKernelMultiDevice(comm->intraParams, comm->intraCudaDevs, comm->intraRanks, *comm->intraCGMode));
      NCCLCHECK(ncclCpuBarrierLast(comm));
    }
  }
  return ncclSuccess;
}

ring allreduce kernel

我们直接看下allreduce kernel，怎么调用到这里的逻辑可以参考第九节，这里不再赘述，

template<int UNROLL, class FUNC, typename T>
__device__ void ncclAllReduceRingKernel(struct CollectiveArgs* args) {
  const int tid = threadIdx.x;
  const int nthreads = args->coll.nThreads-WARP_SIZE;
  const int bid = args->coll.bid;
  const int nChannels = args->coll.nChannels;
  struct ncclDevComm* comm = args->comm;
  struct ncclChannel* channel = comm->channels+blockIdx.x;
  struct ncclRing* ring = &channel->ring;
  const int stepSize = comm->buffSizes[NCCL_PROTO_SIMPLE] / (sizeof(T)*NCCL_STEPS);
  const int chunkSize = stepSize * ALLREDUCE_CHUNKSTEPS;
  const int nranks = comm->nRanks;
  const ssize_t loopSize = nChannels*(ssize_t)chunkSize;
  const ssize_t size = args->coll.count;
  // Compute pointers
  const T * __restrict__ thisInput = (const T*)args->sendbuff;
  T * __restrict__ thisOutput = (T*)args->recvbuff;

  ncclPrimitives<UNROLL, ALLREDUCE_CHUNKSTEPS/ALLREDUCE_SLICESTEPS, ALLREDUCE_SLICESTEPS, T, 1, 1, 1, FUNC>
    prims(tid, nthreads, &ring->prev, &ring->next, thisOutput, stepSize, channel, comm);

  ...
}

类似send/recv kernel，allreduce kernel也会有一个专门的warp做sync以降低延迟。

然后这里会看到几个概念，step，slice和chunk，buffer会被切分为NCCL_STEPS个slot，一个slot就是一个step，因此stepSize就是bufferSize / (sizeof(T) * NCCL_STEPS)；如图二提到的，一个rank一次收发比如1M的数据，这个1M数据就是一个chunk，通信原语ncclPrimitives的api比如directSend，一次传输的就是一个chunk的数据；一个chunk有多个step，即ALLREDUCE_CHUNKSTEPS，所以chunkSize就是stepSize * ALLREDUCE_CHUNKSTEPS；在directSend内部，会将chunk切分为多个slice，一个slice也是多个step，prmitives里数据通信和同步的实际粒度为slice。

一个kernel一共有nChannels个block，所以一次循环过程中一个rank会处理loopSize = nChannels * chunkSize长度的数据。sendbuff为用户传入的输入数据，recvbuff为用户传入的输出数据，然后初始化ncclPrimitives，SLICESPERCHUNK为一个chunk等于几个slice，slicesteps为一个slice是几个step，NRECV表示从几个地方收数据，NSEND表示发送给几个地方，ring allreduce的NRECV和NSEND均为1，DIRECT表示是否支持直接收发，这个下边会介绍。构造函数中recvPeers就是从哪里接收，长度为NRECV，就是ring中的前一个rank，sendPeers同理。

template <int UNROLL, int SLICESPERCHUNK, int SLICESTEPS, typename T, int NRECV, int NSEND, int DIRECT, class FUNC>
class ncclPrimitives {
	...
	public:
  __device__ __forceinline__
  ncclPrimitives(const int tid, const int nthreads, int* recvPeers, int* sendPeers, T* directBuff, int stepSize, struct ncclChannel* channel, struct ncclDevComm* comm)
    : comm(comm), tid(tid), nthreads(nthreads), wid(tid%WARP_SIZE), stepSize(stepSize) {
    	...
    }
    ...
}

template<int UNROLL, class FUNC, typename T>
__device__ void ncclAllReduceRingKernel(struct CollectiveArgs* args) {
  ...
  for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += nranks*loopSize) {
    ssize_t realChunkSize = min(chunkSize, DIVUP(size-gridOffset,nranks*nChannels));
    ALIGN_SIZE(realChunkSize, nthreads*sizeof(uint64_t)/sizeof(T));
    ssize_t chunkOffset = gridOffset + bid*nranks*realChunkSize;

    /// begin AllReduce steps ///
    ssize_t offset;
    int nelem;
    int chunk;

    // step 0: push data to next GPU
    chunk = ring->devUserRanks[nranks-1];
    offset = chunkOffset + chunk * realChunkSize;
    nelem = min(realChunkSize, size-offset);

    prims.send(thisInput+offset, nelem);
    ...
}

按照图二的例子的话，一个block一次处理图二中的1M数据，例如a0，所以4个rank加起来能处理的总长度就是4M，即(a0+a1+a2+a3)，第二个block从a5开始处理，因此4个rank所有block一次循环能处理nranks*loopSize长度的数据，因此gridOffset每次加这么多。

然后开始执行reduce scatter的第一步，将数据从用户的输入thisInput发送到下一个rank的buffer。devUserRanks是以当前rank为起点按序保存了当前ring的所有rank，这里和图二不同的一点是第一次发送的是当前ring最后一个rank的数据，比如rank0发送的是a3，后续按照实际代码来讲。然后执行prims.send将a3发送到下一个rank的buffer里。

template<int UNROLL, class FUNC, typename T>
__device__ void ncclAllReduceRingKernel(struct CollectiveArgs* args) {
  ...

  for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += nranks*loopSize) {
	...
    // k-2 steps: reduce and copy to next GPU
    for (int j=2; j<nranks; ++j) {
      chunk = ring->devUserRanks[nranks-j];
      offset = chunkOffset + chunk * realChunkSize;
      nelem = min(realChunkSize, size-offset);

      prims.recvReduceSend(thisInput+offset, nelem);
    }

    // step k-1: reduce this buffer and data, which will produce the final
    // result that we store in this data and push to the next GPU
    chunk = ring->devUserRanks[0];
    offset = chunkOffset + chunk * realChunkSize;
    nelem = min(realChunkSize, size-offset);

    prims.directRecvReduceCopySend(thisInput+offset, thisOutput+offset, offset, nelem);

    ...
  }
}

然后继续执行reduce scatter过程的nranks - 2步，每次都通过recvReduceSend将自己buffer中已经接收到的数据和thisInput中的数据进行reduce，比如求和，然后将结果发送给next rank的buffer。以rank0为例，最后rank0通过directRecvReduceCopySend将a0和prev rank发送过来的数据进行reduce，然后发送给next rank的buffer和thisOutput，此时reduce scatter执行完成，每个rank都拿到了一块完整的数据，例如对于rank0的完整数据就是a0对应的数据，并将自己对应的完整数据拷贝到了下一个rank的buffer和用户api输入的recvbuff中。

template<int UNROLL, class FUNC, typename T>
__device__ void ncclAllReduceRingKernel(struct CollectiveArgs* args) {
  ...

  for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += nranks*loopSize) {
    ...
    // k-2 steps: copy to next GPU
    for (int j=1; j<nranks-1; ++j) {
      chunk = ring->devUserRanks[nranks-j];
      offset = chunkOffset + chunk * realChunkSize;
      nelem = min(realChunkSize, size-offset);

      prims.directRecvCopySend(thisOutput+offset, offset, nelem);
    }

    // Make final copy from buffer to dest.
    chunk = ring->devUserRanks[1];
    offset = chunkOffset + chunk * realChunkSize;
    nelem = min(realChunkSize, size-offset);

    // Final wait/copy.
    prims.directRecv(thisOutput+offset, offset, nelem);
  }
}

然后开始执行allgather，先执行nranks - 2步的directRecvCopySend，将当前buffer里nelem长度的数据发送给next rank的buffer以及用户api输入的recvbuff中。第nranks - 1步通过directRecv将当前buffer里nelem长度的数据发送到用户api输入的recvbuff中，不需要发送给next rank了。

到这里就完成了ring allreduce kernel的执行，然后我们看下ring allreduce kernel中用到的ncclPrimitives api。

ncclPrimitives

第九节介绍过ncclPrimitives如何收发数据和同步的，但是send/recv场景只用到了directSend和directRecv，这里我们主要介绍下ring allreduce场景中各个ncclPrimitives api的作用。

ncclPrimitives api都是使用的GenericOp，GenericOp核心作用就是根据参数创建srcs和dsts数组，然后将srcs数组规约，规约的结果拷贝到dsts中的每一个输出。

template <int DIRECTRECV, int DIRECTSEND, int RECV, int SEND, int SRC, int DST>
  inline __device__ void
  GenericOp(const T* srcPtr, T* dstPtr, int nelem, ssize_t directOffset) {
    int offset = 0;
    int sliceSize = stepSize*SLICESTEPS;
    int dataSize = max(DIVUP(nelem, 16*SLICESPERCHUNK)*16, sliceSize/32);

    const T* srcs[RECV*NRECV+SRC];
    srcs[0] = SRC ? srcPtr : directRecvPtr<DIRECTRECV>(0, directOffset);
    if (RECV) {
      if (SRC) srcs[1] = recvPtr(0);
      for (int i=1; i<NRECV && i<nrecv; i++) srcs[SRC+i] = recvPtr(i);
    }

    T* dsts[SEND*NSEND+DST];
    dsts[0] = DST ? dstPtr : directSendPtr<DIRECTSEND>(0, directOffset);
    if (SEND) {
      if (DST) dsts[1] = directSendPtr<DIRECTSEND>(0, directOffset);
      for (int i=1; i<NSEND && i<nsend; i++) dsts[DST+i] = directSendPtr<DIRECTSEND>(i, directOffset);
    }
	...
}

模板参数中RECV表示是否需要recv数据，SEND表示是否需要send数据，SRC表示srcs数组中是否有参数srcPtr，如果SRC不为0，那么srcs数组中第一个是srcPtr，第二个是前一个rank的buffer，如果SRC为0，那么srcs只有一个，就是前一个GPU的buffer；DST同理。

DIRECTSEND和DIRECTRECV在这个场景没有什么差异，我们先忽略，稍等介绍一下。

api

send

  __device__ __forceinline__ void
  send(const T* src, int nelem) {
    GenericOp<0, 0, 0, 1, 1, 0>(src, NULL, nelem, 0); 
  }

send接口的SEND和SRC为1，所以srcs数组只有src，dsts数组只有下一个rank的buffer，因此send的作用就是将src里的nelem长度的数据发送给下一个rank的buffer

recvReduceSend

  __device__ __forceinline__ void
  recvReduceSend(const T* src, int nelem) {
    GenericOp<0, 0, 1, 1, 1, 0>(src, NULL, nelem, 0); 
  }

SRC和RECV都为1，因此srcs数组为src和前一个rank的buffer，SEND为1但DST为0，因此dsts数组只有下一个rank的buffer，因此recvReduceSend的作用就是将src里的nelem长度的数据和prev rank发送过来的数据进行reduce，然后发送给next rank的buffer

directRecvReduceCopySend

  __device__ __forceinline__ void
  directRecvReduceCopySend(const T* src, T* dst, ssize_t directOffset, int nelem) {
    // Direct is only for the send part
    GenericOp<0, 1, 1, 1, 1, 1>(src, dst, nelem, directOffset);
  }

SRC和RECV均为1，因此srcs数组为src和当前rank的buffer，SEND和DST均为1，因此dsts数组为dst和下一个rank的buffer，因此directRecvReduceCopySend的作用就是将src里的nlem长度的数据和prev rank发送到当前buffer的数据进行reduce，然后发送给next rank的buffer和dst

directRecvCopySend

  __device__ __forceinline__ void
  directRecvCopySend(T* dst, ssize_t directOffset, int nelem) {
    GenericOp<1, 1, 1, 1, 0, 1>(NULL, dst, nelem, directOffset);
  }

由于SRC为0其他均为1，因此srcs数组只有当前rank的buffer，dsts数组为dst和下一个rank的buffer，因此directRecvCopySend的作用就是将当前rank buffer里收到的数据拷贝到dst和下一个rank的buffer

directRecv

  __device__ __forceinline__ void
  directRecv(T* dst, ssize_t directOffset, int nelem) {
    GenericOp<1, 0, 1, 0, 0, 1>(NULL, dst, nelem, directOffset);
  }

由于RECV为1，SRC为0，因此srcs只有当前rank的buffer，DST为1，因此dsts数组为dst，因此directRecv的作用就是将数据从当前rank的buffer里收到的数据拷贝到dst

direct

然后介绍下之前提到的direct的作用，比如rank0执行send，rank1执行recv，那么rank0将数据从src拷贝到rank1的buffer，rank1将数据从buffer拷贝到dst，而如果是directSend，那么有可能会bypass rank1的buffer，直接发送给dst。为什么是有可能，以及rank0如何知道dst是哪里，我们回顾下transport的建立。

struct ncclSendMem {
  union {
    struct {
      uint64_t head;
      char pad1[CACHE_LINE_SIZE-sizeof(uint64_t)];
      void* ptrExchange;
      char pad2[CACHE_LINE_SIZE-sizeof(void*)];
    };  
    char pad3[MEM_ALIGN];
  };  
  char buff[1]; // Actually larger than that
};

send端有个变量叫ptrExchange。

static ncclResult_t p2pSendConnect(struct ncclConnect* connectInfo, int nranks, int rank, struct ncclConnector* send) {
  ...
  send->conn.ptrExchange = &resources->devMem->ptrExchange;
}

ncclResult_t p2pRecvConnect(struct ncclConnect* connectInfo, int nranks, int rank, struct ncclConnector* recv) {
  ...
  if (info->direct) {
    remDevMem = (struct ncclSendMem*)(info->directPtr);
    if (info->read == 0) {
      recv->conn.direct |= NCCL_DIRECT_GPU;
      recv->conn.ptrExchange = &remDevMem->ptrExchange;
    }   
  }
  ...
}

send端和recv端connect的过程中，recv端会保存下send端的ptrExchange。

  __device__ __forceinline__ void loadSendConn(struct ncclConnInfo* conn, int i) {
    sendBuff[i] = (T*)conn->buffs[NCCL_PROTO_SIMPLE];
   	...
    if (DIRECT && (conn->direct & NCCL_DIRECT_GPU)) {
      void* volatile* ptr = conn->ptrExchange;
      while ((sendDirectBuff[i] = (T*)(*ptr)) == NULL);
      barrier();
      if (tid == 0) *ptr = NULL;
    }   
    ...
  }
  __device__ __forceinline__ void loadRecvConn(struct ncclConnInfo* conn, int i, T* directBuff) {
    recvBuff[i] = (const T*)conn->buffs[NCCL_PROTO_SIMPLE];
    ...
    if (DIRECT && (conn->direct & NCCL_DIRECT_GPU)) {
      recvDirectBuff[i] = directBuff;
      if (tid == 0) *conn->ptrExchange = directBuff;
    }
    ...
  }

然后ncclPrimitives加载conn的时候，recv端会将directBuff，也就是dst，写入到ptrExchange，所以send端就知道了dst在哪里。

同时也能看到限制，只有info->direct == 1且info->read == 0的时候才支持direct，也就是说必须为同一个进程并且使用p2p write才支持，我们这个场景使用了p2p read，因此直接忽略了direct。

多机

多机的实际通信流程和第十节的多机send/recv一致，我们主要关注下不一致的地方。

初始化

假设单机内执行ncclTopoCompute搜索出的ring为：

NET/0 GPU/0 GPU/1 GPU/2 GPU/3 GPU/4 GPU/5 GPU/6 GPU/7 NET/0

相比单机的环，变成了一条链，前后加了网卡。建链的过程只是多了机器间的建链，不再赘述。

ncclSaveKernel

ncclResult_t ncclSaveKernel(struct ncclInfo* info) {
  ...
  struct ncclColl coll;
  struct ncclProxyArgs proxyArgs;
  memset(&proxyArgs, 0, sizeof(struct ncclProxyArgs));
  NCCLCHECK(computeColl(info, &coll, &proxyArgs));
  ...
  for (int bid=0; bid<nChannels*nSubChannels; bid++) {
    ...
    // Proxy
    proxyArgs.channel = channel;
    // Adjust pattern for CollNet based on channel index
    if (nSubChannels == 2) {
      info->pattern = (channelId < info->comm->nChannels/nSubChannels) ? ncclPatternCollTreeUp : ncclPatternCollTreeDown;
    }

    if (info->coll == ncclCollSendRecv) {
      info->comm->myParams->gridDim.x = std::max<unsigned>(info->comm->myParams->gridDim.x, channelId+1);
      NCCLCHECK(ncclProxySaveP2p(info, channel));
    } else {
      NCCLCHECK(ncclProxySaveColl(&proxyArgs, info->pattern, info->root, info->comm->nRanks));
    }
    ...
  }
  info->comm->opCount++;
  return ncclSuccess;
}

ncclSaveKernel中唯一有区别的就是computeColl和ncclProxySaveColl。

computeColl

  static ncclResult_t computeColl(struct ncclInfo* info /* input */, struct ncclColl* coll, struct ncclProxyArgs* proxyArgs /* output */) {
  ...
  int stepSize   = info->comm->buffSizes[info->protocol]/NCCL_STEPS;
  int chunkSteps = (info->protocol == NCCL_PROTO_SIMPLE && info->algorithm == NCCL_ALGO_RING) ? info->chunkSteps : 1;
  int sliceSteps = (info->protocol == NCCL_PROTO_SIMPLE && info->algorithm == NCCL_ALGO_RING) ? info->sliceSteps : 1;
  int chunkSize  = stepSize*chunkSteps;
  ...
  // Compute nSteps for proxies
  int chunkEffectiveSize = chunkSize;
  if (info->protocol == NCCL_PROTO_LL) chunkEffectiveSize /= 2;
  if (info->protocol == NCCL_PROTO_LL128) chunkEffectiveSize = (chunkSize / NCCL_LL128_LINEELEMS) * NCCL_LL128_DATAELEMS;
  int nLoops = (int)(DIVUP(info->nBytes, (((size_t)(info->nChannels))*info->nchunksPerLoop*chunkEffectiveSize)));
  proxyArgs->nsteps = info->nstepsPerLoop * nLoops * chunkSteps;
  proxyArgs->sliceSteps = sliceSteps;
  proxyArgs->chunkSteps = chunkSteps;
  proxyArgs->protocol = info->protocol;
  proxyArgs->opCount = info->comm->opCount;
  proxyArgs->dtype = info->datatype;
  proxyArgs->redOp = info->op;
  ...
  return ncclSuccess;
}

nLoops表示一共有多少次循环，如图二描述，一个chunk就是a0，大小为chunkEffectiveSize，一次执行能处理nchunksPerLoop*chunkEffectiveSize的数据，即(a0+a1+a2+a3)，一共有nChannels，所以一次处理的数据再乘上nChannels，然后去除nBytes就算出了一共需要循环执行多少次。

然后计算nsteps，表示一共有多少个step，step就是buff中的一个slot，因为处理一个chunk的时候会发送数据nstepsPerLoop次，一个chunk里有chunkSteps个step，一共有nLoops个循环，因此nsteps就是nstepsPerLoop * nLoops * chunkSteps。

所以proxy就通过proxyArgs里的这些信息知道了应该发送多少次数据。

ncclProxySaveColl

然后将每个channel的proxyArgs加到comm的args链表中

ncclResult_t ncclProxySaveColl(struct ncclProxyArgs* args, int pattern, int root, int nranks) {
  if (pattern == ncclPatternRing || pattern == ncclPatternRingTwice || pattern == ncclPatternPipelineFrom || pattern == ncclPatternPipelineTo) {
    struct ncclRing* ring = &args->channel->ring;
    if (NeedProxy(RECV, pattern, root, ring, nranks)) NCCLCHECK(SaveProxy<proxyRecv>(ring->prev, args));
    if (NeedProxy(SEND, pattern, root, ring, nranks)) NCCLCHECK(SaveProxy<proxySend>(ring->next, args));
  }
  ...
}

NeedProxy固定返回true，然后执行SaveProxy

template <int type>
static ncclResult_t SaveProxy(int peer, struct ncclProxyArgs* args) {
  if (peer < 0) return ncclSuccess;

  struct ncclPeer* peerComm = args->channel->peers+peer;
  struct ncclConnector* connector = type == proxyRecv ? &peerComm->recv : &peerComm->send;
  if (connector->transportComm == NULL) {
    WARN("[%d] Error no transport for %s peer %d on channel %d\n", connector->comm->rank,
        type == proxyRecv ? "recv" : "send", peer, args->channel->id);
    return ncclInternalError;
  }
  if (connector->transportComm->proxy == NULL) return ncclSuccess;

  struct ncclProxyArgs* op; 
  NCCLCHECK(allocateArgs(connector->comm, &op));
  memcpy(op, args, sizeof(struct ncclProxyArgs));
  op->connector = connector;
  op->progress = connector->transportComm->proxy;
  op->state = ncclProxyOpReady;
  ProxyAppend(connector, op);
  return ncclSuccess;
}

因为建链接的过程中只有rank7的send是netTransport，所以只有rank7的send会执行ProxyAppend，同理只有rank0的recv会执行ProxyAppend。

数据发送

ProxyAppend不再赘述，就是将args加入到comm中的链表，launch kernel之后就会唤醒proxy线程，proxy遍历这个链表，执行对应的操作，以send为例。

ncclResult_t netSendProxy(struct ncclProxyArgs* args) {
  struct netSendResources* resources = (struct netSendResources*) (args->connector->transportResources);
  if (args->state == ncclProxyOpReady) {
    // Round to next multiple of sliceSteps
    resources->step = ROUNDUP(resources->step, args->chunkSteps);
    args->head = resources->step;
    args->tail = resources->step;
    args->end = args->head + args->nsteps;
    args->state = ncclProxyOpProgress;
  }
  if (args->state == ncclProxyOpProgress) {
    int p = args->protocol;
    int stepSize = args->connector->comm->buffSizes[p] / NCCL_STEPS;
    char* localBuff = args->connector->conn.buffs[p];
    void* mhandle = *(resources->mhandlesProto[p]);
    args->idle = 1;
    if (args->head < args->end) {
      int buffSlot = args->tail%NCCL_STEPS;
      if (args->tail < args->end && args->tail < args->head + NCCL_STEPS) {
        volatile int* sizesFifo = resources->recvMem->sizesFifo;
        volatile uint64_t* recvTail = &resources->recvMem->tail;
 		...
 		else if (args->tail < *recvTail) {
          // Send through network
          if (sizesFifo[buffSlot] != -1) {
            NCCLCHECK(ncclNetIsend(resources->netSendComm, localBuff+buffSlot*stepSize, sizesFifo[buffSlot], mhandle, args->requests+buffSlot));
            if (args->requests[buffSlot] != NULL) {
              sizesFifo[buffSlot] = -1;
              // Make sure size is reset to zero before we update the head.
              __sync_synchronize();
              args->tail += args->sliceSteps;
              args->idle = 0;
            }
          }
        }
      }
      if (args->head < args->tail) {
        int done;
        int buffSlot = args->head%NCCL_STEPS;
        NCCLCHECK(ncclNetTest(args->requests[buffSlot], &done, NULL));
        if (done) {
          args->head += args->sliceSteps;
          resources->sendMem->head = args->head;
          args->idle = 0;
        }
      }
    }
    if (args->head == args->end) {
      resources->step = args->end;
      args->idle = 0;
      args->state = ncclProxyOpNone;
    }
  }
  return ncclSuccess;
}

可以看到end就是通过nsteps计算得到的，所以proxy知道整个算法流程一共需要多少个slot，因为primitives一次实际发送sliceSteps个slot，因此每次对head和tail的移动都是sliceSteps。