NCCL源码解析⑥：Channel搜索

作者｜KIDGINBROOK
更新｜潘丽晨

上节讲到已经计算出GPU和NIC节点到其他任意节点的最优路径了，本节看下NCCL中channel的搜索过程。

NCCL中channel的概念表示一个通信路径，为了更好地利用带宽和网卡，以及同一块数据可以通过多个channel并发通信，另外后续可以看到一个channel对应了一个GPU SM，所以基于这些原因，NCCL会使用多channel，搜索的过程就是搜索出来一组channel。

如上节所述，单机的情况下会在ncclTopoTrimSystem函数里删除网卡，因此我们先看下单机八卡这种简化的情况，最后再看下多机引入网卡之后的情况。

static float getMaxWidth(struct ncclTopoSystem* system, struct ncclTopoNode* gpu, int type) {
  float maxWidth = 0.0;
  for (int i=0; i<system->nodes[type].count; i++) {
    struct ncclTopoLinkList* path = gpu->paths[type]+i;
    float width = path->width;
    if (path->count == 0) continue;
    maxWidth = std::max(maxWidth, width);
  }
  return maxWidth;
}
 
 
ncclResult_t ncclTopoSearchInit(struct ncclTopoSystem* system) {
  system->maxWidth = 0.0;
  int inter = system->nodes[NET].count;
  if (inter == 0 && system->nodes[GPU].count == 1) {
    system->maxWidth = LOC_WIDTH;
    return ncclSuccess;
  }
  for (int g=0; g<system->nodes[GPU].count; g++) {
    struct ncclTopoNode* gpu = system->nodes[GPU].nodes+g;
    system->maxWidth = std::max(system->maxWidth, getMaxWidth(system, gpu, inter ? NET : GPU));
  }
  return ncclSuccess;
}

ncclTopoSearchInit就是初始化system->maxWidth，如果是单机单卡的情况，那么maxWidth设置为LOC_WIDTH，否则就遍历每个GPU节点，查看到其他所有GPU节点或者网卡最大带宽。

struct ncclTopoGraph ringGraph;
  ringGraph.id = 0;
  ringGraph.pattern = NCCL_TOPO_PATTERN_RING;
  ringGraph.crossNic = ncclParamCrossNic();
  ringGraph.collNet = 0;
  ringGraph.minChannels = 1;
  ringGraph.maxChannels = MAXCHANNELS/2;
  NCCLCHECK(ncclTopoCompute(comm->topo, &ringGraph));
  NCCLCHECK(ncclTopoPrintGraph(comm->topo, &ringGraph));

nccl执行集合通信时支持ring，tree和collnet三种算法，这里我们以ring来举例，后续专门介绍ring和tree。

struct ncclTopoGraph {
  // Input / output
  int id; // ring : 0, tree : 1, collnet : 2
  int pattern;
  int crossNic;
  int collNet;
  int minChannels;
  int maxChannels;
  // Output
  int nChannels;      // 搜索到的channel数量
  float speedIntra;   // 节点内单个channel带宽
  float speedInter;   // 节点间单个channel带宽
  int typeIntra;      // 节点内channel的路径类型
  int typeInter;      // 节点间channel的路径类型
  int sameChannels;   // channel是否一样
  int nHops;
  int intra[MAXCHANNELS*NCCL_TOPO_MAX_NODES];  // 节点内每个channel路径
  int inter[MAXCHANNELS*2];                    // 节点间每个channel路径
};

ncclTopoGraph记录了搜索到的结果，具体含义见注释。

然后看下ncclTopoCompute，这里就是实际搜索channel的过程，目标是搜索出来尽可能多，带宽尽可能大的一系列channel，本质就是暴力搜索，先设置一系列的条件搜答案，如果搜不出来则降低条件继续搜。

由于此时没有NET节点，所以crossNic为0，然后初始化graph，首先设置最高的条件，限制节点内部只能使用不超过PATH_NVL路径，节点间只能使用不超过PATH_PIX的路径，然后通过system-maxWidth设置speedIntra和speedInter，接着执行ncclTopoSearchRec搜索出一个答案存储到tmpGraph中。

如果此时就是最优的结果，channel数等于maxChannel，并且speedInter也等于maxWidth，则直接退出，否则就开始逐步降低条件，比如将sameChannel设置为0，允许channel之间不一样；调大typeIntra和typeInter；允许crossNic；调小speedInter和speedIntra。

ncclResult_t ncclTopoCompute(ncclTopoSystem* system, struct ncclTopoGraph* graph) {
  int ngpus = system->nodes[GPU].count;
  int crossNic = (system->nodes[NET].count > 1) && graph->crossNic ? 1 : 0;
  graph->speedIntra = graph->speedInter = 0;
  if (graph->crossNic == 2) graph->crossNic = 0;
  graph->typeIntra = ngpus == 1 ? PATH_LOC : PATH_NVL;
  graph->typeInter = PATH_PIX;
  graph->nChannels = 0;
  graph->sameChannels = 1;
 
 
  if (ngpus == 1) if (graph->pattern != NCCL_TOPO_PATTERN_RING) graph->pattern = NCCL_TOPO_PATTERN_TREE;
 
  struct ncclTopoGraph tmpGraph;
  memcpy(&tmpGraph, graph, sizeof(struct ncclTopoGraph));
 
  // First try crossnic, then decrease speed and finally increase speedIntra.
  tmpGraph.pattern = graph->pattern;
  int pass = 1;
  int speedIndex = 0;
  while (speedArray[speedIndex] > system->maxWidth && speedIndex < NSPEEDS-1) speedIndex++;
  tmpGraph.speedIntra = tmpGraph.speedInter = speedArray[speedIndex];
  int64_t globalTimeout = NCCL_SEARCH_GLOBAL_TIMEOUT;
 
search:
  int time = tmpGraph.sameChannels ? NCCL_SEARCH_TIMEOUT_SAMECHANNELS :
    tmpGraph.pattern == NCCL_TOPO_PATTERN_TREE ? NCCL_SEARCH_TIMEOUT_TREE : NCCL_SEARCH_TIMEOUT;
  tmpGraph.nChannels = 0;
  globalTimeout -= time;
 
  NCCLCHECK(ncclTopoSearchRec(system, &tmpGraph, graph, &time));
 
  // Optimal solution, stop here
  if (graph->nChannels == graph->maxChannels && graph->speedInter == system->maxWidth) goto done;
 
  if (pass == 1) {
    // First pass, we don't have a solution yet ; try other options
 
    // Try having different channels
    if (tmpGraph.sameChannels == 1) {
      tmpGraph.sameChannels = 0;
      goto search;
    }
    tmpGraph.sameChannels = 1;
 
    if (time != -1) globalTimeout += time;
    else globalTimeout = NCCL_SEARCH_GLOBAL_TIMEOUT;
    if (globalTimeout < 0) goto done;
 
    int maxTypeIntra = system->nodes[NET].count > 0 ? tmpGraph.typeInter : PATH_SYS;
    if (tmpGraph.typeIntra < maxTypeIntra && (graph->nChannels == 0 || tmpGraph.typeIntra < graph->typeIntra)) {
      tmpGraph.typeIntra += 1;
      goto search;
    }
    tmpGraph.typeIntra = ngpus == 1 ? PATH_LOC : PATH_NVL;
    if (system->nodes[NET].count > 0 && tmpGraph.typeInter < PATH_SYS && (graph->nChannels == 0 || tmpGraph.typeInter < graph->typeInter || tmpGraph.typeInter < PATH_PXB)) {
      tmpGraph.typeInter += 1;
      goto search;
    }
    tmpGraph.typeInter = PATH_PIX;
 
    // Try a simpler tree
    if (tmpGraph.pattern == NCCL_TOPO_PATTERN_SPLIT_TREE_LOOP) {
      tmpGraph.pattern = NCCL_TOPO_PATTERN_SPLIT_TREE;
      goto search;
    }
    if (tmpGraph.pattern == NCCL_TOPO_PATTERN_SPLIT_TREE) {
      tmpGraph.pattern = NCCL_TOPO_PATTERN_TREE;
      goto search;
    }
    tmpGraph.pattern = graph->pattern;
 
    if (crossNic && tmpGraph.crossNic == 0) {
      // Try again with crossNic if permitted
      tmpGraph.crossNic = crossNic;
      goto search;
    }
    tmpGraph.crossNic = 0;
 
    // Decrease speed until we find a solution
    if ((speedIndex < NSPEEDS-1) && (graph->nChannels == 0 || (speedArray[speedIndex+1]/graph->speedInter > .49))) {
      tmpGraph.speedInter = tmpGraph.speedIntra = speedArray[++speedIndex];
      goto search;
    }
    speedIndex = 0;
    while (speedArray[speedIndex] > system->maxWidth && speedIndex < NSPEEDS-1) speedIndex++;
    tmpGraph.speedIntra = tmpGraph.speedInter = speedArray[speedIndex];
 
  }
 
done:
  // We have a solution. Start from that solution and move to pass 2.
  if (pass == 1) {
    time = -1;
    memcpy(&tmpGraph, graph, sizeof(tmpGraph));
    speedIndex = 0;
    while (speedArray[speedIndex] > graph->speedInter && speedIndex < NSPEEDS-1) speedIndex++;
    tmpGraph.speedIntra = tmpGraph.speedInter = speedArray[speedIndex];
    tmpGraph.minChannels = graph->nChannels;
    pass = 2;
  }
 
  // 3. See if we can increase speedIntra for trees (2 nodes or collnet)
  if (pass == 2) {
    if (time != 0 && graph->pattern != NCCL_TOPO_PATTERN_RING &&
        tmpGraph.speedIntra == graph->speedIntra && tmpGraph.speedIntra < tmpGraph.speedInter*2 &&
        speedIndex > 0) {
      tmpGraph.speedIntra = speedArray[--speedIndex];
      goto search;
    }
    time = -1;
    memcpy(&tmpGraph, graph, sizeof(tmpGraph));
  }
 
  if (graph->nChannels == 0 && graph->collNet == 0) {
    WARN("Could not find a path for pattern %d, falling back to simple order\n", graph->pattern);
    for (int i=0; i<ngpus; i++) graph->intra[i] = system->nodes[GPU].nodes[i].gpu.rank;
    graph->inter[0] = graph->inter[1] = 0;
    graph->speedIntra = graph->speedInter = 0.1;
    graph->typeIntra = graph->typeInter = PATH_SYS;
    graph->nChannels = 1;
  }
 
  if (graph->speedIntra >= 25.0) {
    int dupChannels = std::min(graph->nChannels*2, graph->maxChannels);
    memcpy(graph->intra+graph->nChannels*ngpus, graph->intra, (dupChannels-graph->nChannels)*ngpus*sizeof(int));
    memcpy(graph->inter+graph->nChannels*2,graph->inter, (dupChannels-graph->nChannels)*2*sizeof(int));
    graph->speedIntra /= DIVUP(dupChannels, graph->nChannels);
    graph->speedInter /= DIVUP(dupChannels, graph->nChannels);
    graph->nChannels = dupChannels;
  }
  return ncclSuccess;
}

然后开始搜索channel，对于ringGraph来说其实就是搜索出来一系列的环，每个rank对应这个环的一个节点，记录了环的prev和next，这里是一个回溯的过程，执行一次ncclTopoSearchRec就会得到一个环，执行一次ncclTopoSearchTryGpu看选择出来的下一个点能不能到达，执行一次ncclTopoSearchRecGpu用来找下一个GPU，接下来具体看下。

ncclResult_t ncclTopoSearchRec(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, struct ncclTopoGraph* saveGraph, int* time) {
  int backToNet, backToFirstRank;
  NCCLCHECK(ncclTopoSearchParams(system, graph->pattern, &backToNet, &backToFirstRank));
  if (system->nodes[NET].count) {
    // Start from NET
    ncclTopoSearchRecNet(system, graph, saveGraph, backToNet, backToFirstRank, time);
  } else {
    // Intra-node only.
    if (graph->nChannels == 0) {
      // Try PCI order first
      NCCLCHECK(ncclTopoSearchTryGpu(system, graph, saveGraph, 0, backToNet, backToFirstRank, FORCED_ORDER_PCI, time, -1, -1, 0));
    } else {
      // Also try to replay previous channel
      int g;
      NCCLCHECK(ncclTopoReplayGetGpu(system, graph, -1, &g));
      NCCLCHECK(ncclTopoSearchTryGpu(system, graph, saveGraph, 0, backToNet, backToFirstRank, FORCED_ORDER_REPLAY, time, -1, -1, g));
    }
    if (graph->sameChannels == 0 || graph->nChannels == 0) {
      // Finally, try all other possibilities unless we are forced to use the same channels
      for (int g=0; g<system->nodes[GPU].count; g++) {
        NCCLCHECK(ncclTopoSearchTryGpu(system, graph, saveGraph, 0, backToNet, backToFirstRank, 0, time, -1, -1, g));
      }
    }
  }
  return ncclSuccess;
}

通过ncclTopoSearchParams设置backToNet和backToFirstRank参数，单机八卡ringGraph场景下这俩参数会分别设置为-1和7，此时nchannel为0，执行ncclTopoSearchTryGpu，强制为pci顺序，就是devid的顺序，从dev0开始。

ncclResult_t ncclTopoSearchParams(struct ncclTopoSystem* system, int pattern, int* backToNet, int* backToFirstRank) {
  if (system->nodes[NET].count) {
    if (pattern == NCCL_TOPO_PATTERN_RING) *backToNet = system->nodes[GPU].count-1;
    else if (pattern == NCCL_TOPO_PATTERN_TREE) *backToNet = 0;
    else *backToNet = 1;
    if (pattern == NCCL_TOPO_PATTERN_SPLIT_TREE_LOOP) *backToFirstRank = system->nodes[GPU].count-1;
    else *backToFirstRank = -1;
  } else {
    *backToNet = -1;
    if (pattern == NCCL_TOPO_PATTERN_RING || pattern == NCCL_TOPO_PATTERN_SPLIT_TREE_LOOP) *backToFirstRank = system->nodes[GPU].count-1;
    else *backToFirstRank = -1;
  }
  return ncclSuccess;
}

然后执行ncclTopoSearchTryGpu，这里会判断下一个点能不能到达，因为type为-1，ncclTopoFollowPath会设置gpu为0号卡，直接执行ncclTopoSearchRecGpu，从0号卡开始搜，step为0。

ncclResult_t ncclTopoSearchTryGpu(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, struct ncclTopoGraph* saveGraph, int step, int backToNet, int backToFirstRank, int forcedOrder, int *time, int type, int index, int g) {
  const uint64_t flag = 1ULL<<(graph->nChannels);
  struct ncclTopoNode* gpu;
  NCCLCHECK(ncclTopoFollowPath(system, graph, type, index, GPU, g, 1, &gpu));
  if (gpu) {
    gpu->used ^= flag;
    NCCLCHECK(ncclTopoSearchRecGpu(system, graph, saveGraph, gpu, step, backToNet, backToFirstRank, forcedOrder, time));
    gpu->used ^= flag;
    NCCLCHECK(ncclTopoFollowPath(system, graph, type, index, GPU, g, -1, &gpu));
  }
  return ncclSuccess;
}

ncclResult_t ncclTopoSearchRecGpu(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, struct ncclTopoGraph* saveGraph, struct ncclTopoNode* gpu, int step, int backToNet, int backToFirstRank, int forcedOrder, int *time) {
  if ((*time) <= 0) return ncclSuccess;
  (*time)--;
 
  int ngpus = system->nodes[GPU].count;
  if (step == ngpus) {
    ...
  }
  graph->intra[graph->nChannels*ngpus+step] = gpu->gpu.rank;
  int g = gpu - system->nodes[GPU].nodes;
  if (step == backToNet) {
    ...
  } else if (step < system->nodes[GPU].count-1) {
    // Go to next GPU
    int next[NCCL_TOPO_MAX_NODES];
    int count;
    if (forcedOrder == FORCED_ORDER_PCI) { // Try the PCI order
      next[0] = step+1;
      count = 1;
    } else if (forcedOrder == FORCED_ORDER_REPLAY) { // Try last channel order
      NCCLCHECK(ncclTopoReplayGetGpu(system, graph, step, next));
      count = 1;
    } else { // Normal search
      NCCLCHECK(ncclTopoSearchNextGpuSort(system, graph, gpu, next, &count, backToNet == -1 ? 0 : backToNet == step+1 ? 1 : -1 ));
    }
    for (int i=0; i<count; i++) {
      NCCLCHECK(ncclTopoSearchTryGpu(system, graph, saveGraph, step+1, backToNet, backToFirstRank, forcedOrder, time, GPU, g, next[i]));
    }
  } else if (step == backToFirstRank) {
    ...
  } else {
    // Next path
    NCCLCHECK(ncclTopoSearchRecGpu(system, graph, saveGraph, gpu, ngpus, -1, -1, forcedOrder, time));
  }
  return ncclSuccess;
}

然后看下ncclTopoSearchRecGpu，这里会选择下一个节点，先将0号卡节点写入到graph->intra的对应位置；由于当前step是0，因此会在xx行选择下一个GPU，next数组表示候选的GPU节点，由于forcedOrder == FORCED_ORDER_PCI，所以候选只有一个，即1号卡，然后对所有候选执行ncclTopoSearchTryGpu判断这一步是否可行并继续选择下一个节点。

然后回到ncclTopoSearchRec开始尝试判断是否可达1号卡，看下ncclTopoFollowPath，这个函数就是判断能否从type1的index1节点到达type2的index2节点，这里可以看到之前在选起点的时候type1为-1，因此直接将node设置为type2的index2就返回；这次我们要判断gpu0到gpu1是否可达，获取index1到index2的路径path，如果index1和index2的类型都是GPU那么speed就设置为graph->speedIntra，即搜索之前设置的条件，mult是函数的入参，表示需要在path上加还是减去speed，向下搜环的时候需要在path上减去speed，当回溯回去的时候需要将speed加回去，然后判断path的type是否大于之前设置的type，即graph->typeIntra，大于的话说明不可达，然后通过followPath将path上的边全都减去speed，如果有边剩下的带宽不够speed，那么通过rewind加回去，此时路径不可达；如果足够的话，则设置node为index2。

static ncclResult_t ncclTopoFollowPath(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, int type1, int index1, int type2, int index2, int mult, struct ncclTopoNode** node) {
  // First handle easy cases
  *node = system->nodes[type2].nodes+index2;
  if (type1 == -1) return ncclSuccess;
  struct ncclTopoNode* node1 = system->nodes[type1].nodes+index1;
  struct ncclTopoLinkList* path = node1->paths[type2]+index2;
  if (path->count == 0 ) return ncclSuccess;
 
  // Now check link type
  *node = NULL;
  int intra = type1 == GPU && type2 == GPU;
  float speed = intra ? graph->speedIntra : graph->speedInter;
  int type = intra ? graph->typeIntra : graph->typeInter;
 
  if (mult == 1 && (path->type > type)) return ncclSuccess;
 
  speed *= mult;
 
  // Check there is enough bandwidth on paths.
  int step = 0;
  NCCLCHECK(followPath(path, node1, path->count, speed, &step));
  if (step < path->count) goto rewind;
 
  // Enough bandwidth : return destination node.
  graph->nHops += mult*path->count;
  *node = system->nodes[type2].nodes+index2;
  return ncclSuccess;
 
rewind:
  // Not enough bandwidth : rewind and exit.
  NCCLCHECK(followPath(path, node1, step, -speed, &step));
  return ncclSuccess;
}

接着递归执行ncclTopoSearchRecGpu，重复上述过程，直到gpu7，这个时候graph->intra中的第一个环是0,1,2,3,4,5,6,7，此时step为backToFirstRank，然后通过获取第一个gpu，即gpu0，然后继续执行ncclTopoFollowPath判断7到0是否可达，如果可达的话继续递归执行ncclTopoSearchRecGpu，此时step == ngpus，即搜索到了一个环，那会将现有的graph去更新最优的saveGraph，判断标准主要是看总的带宽，即环的数量乘以speedIntra；如果搜到的环的数量已经达到maxChannel了，则结束本次搜索，否则继续递归执行ncclTopoSearchRec搜索下一个环。

ncclResult_t ncclTopoSearchRecGpu(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, struct ncclTopoGraph* saveGraph, struct ncclTopoNode* gpu, int step, int backToNet, int backToFirstRank, int forcedOrder, int *time) {
  if ((*time) <= 0) return ncclSuccess;
  (*time)--;
 
  int ngpus = system->nodes[GPU].count;
  if (step == ngpus) {
    // Determine whether we found a better solution or not
    int copy = 0;
    graph->nChannels++;
    NCCLCHECK(ncclTopoCompareGraphs(graph, saveGraph, ©));
    if (copy) {
      memcpy(saveGraph, graph, sizeof(struct ncclTopoGraph));
      if (graph->nChannels == graph->maxChannels) *time = -1;
    }
    if (graph->nChannels < graph->maxChannels) {
      NCCLCHECK(ncclTopoSearchRec(system, graph, saveGraph, time));
    }
    graph->nChannels--;
    return ncclSuccess;
  }
  graph->intra[graph->nChannels*ngpus+step] = gpu->gpu.rank;
  int g = gpu - system->nodes[GPU].nodes;
  if (step == backToNet) {
    ...
  } else if (step < system->nodes[GPU].count-1) {
    ...
  } else if (step == backToFirstRank) {
    // Find first GPU and loop back to it
    int p;
    NCCLCHECK(getGpuIndex(system, graph->intra[graph->nChannels*ngpus], &p));
    struct ncclTopoNode* firstGpu;
    NCCLCHECK(ncclTopoFollowPath(system, graph, GPU, g, GPU, p, 1, &firstGpu));
    if (firstGpu) {
      NCCLCHECK(ncclTopoSearchRecGpu(system, graph, saveGraph, firstGpu, step+1, backToNet, -1, forcedOrder, time));
      NCCLCHECK(ncclTopoFollowPath(system, graph, GPU, g, GPU, p, -1, &firstGpu));
    }
  } else {
    // Next path
    NCCLCHECK(ncclTopoSearchRecGpu(system, graph, saveGraph, gpu, ngpus, -1, -1, forcedOrder, time));
  }
  return ncclSuccess;
}

假设现在开始搜索下一个环，回到ncclTopoSearchRec，接下来会尝试复制刚刚的环，ncclTopoReplayGetGpu会获取上一个环的第step + 1个gpu，这里其实就是gpu0，然后继续执行ncclTopoSearchTryGpu，这里设置forcedOrder为FORCED_ORDER_REPLAY。

ncclResult_t ncclTopoSearchRec(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, struct ncclTopoGraph* saveGraph, int* time) {
    {
      // Also try to replay previous channel
      int g;
      NCCLCHECK(ncclTopoReplayGetGpu(system, graph, -1, &g));
      NCCLCHECK(ncclTopoSearchTryGpu(system, graph, saveGraph, 0, backToNet, backToFirstRank, FORCED_ORDER_REPLAY, time, -1, -1, g));
    }
}
 
ncclResult_t ncclTopoReplayGetGpu(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, int step, int* g) {
  *g = -1;
  if (graph->nChannels == 0) return ncclInternalError;
  int ngpus = system->nodes[GPU].count;
  int nextRank = graph->intra[(graph->nChannels-1)*ngpus+step+1];
  for (int i=0; i<ngpus; i++) if (system->nodes[GPU].nodes[i].gpu.rank == nextRank) {
    *g = i;
    return ncclSuccess;
  }
  if (*g == -1) return ncclInternalError;
  return ncclSuccess;
}

然后FORCED_ORDER_REPLAY会在寻找下一个节点时通过ncclTopoReplayGetGpu获取上一个环对应step的gpu，因此就是一直在复制上一个环。

ncclResult_t ncclTopoSearchRecGpu(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, struct ncclTopoGraph* saveGraph, struct ncclTopoNode* gpu, int step, int backToNet, int backToFirstRank, int forcedOrder, int *time) {
  ...
  else if (step < system->nodes[GPU].count-1) {
    // Go to next GPU
    int next[NCCL_TOPO_MAX_NODES];
    int count;
    if (forcedOrder == FORCED_ORDER_PCI) { // Try the PCI order
      next[0] = step+1;
      count = 1;
    } else if (forcedOrder == FORCED_ORDER_REPLAY) { // Try last channel order
      NCCLCHECK(ncclTopoReplayGetGpu(system, graph, step, next));
      count = 1;
    } else { // Normal search
      NCCLCHECK(ncclTopoSearchNextGpuSort(system, graph, gpu, next, &count, backToNet == -1 ? 0 : backToNet == step+1 ? 1 : -1 ));
    }
    for (int i=0; i<count; i++) {
      NCCLCHECK(ncclTopoSearchTryGpu(system, graph, saveGraph, step+1, backToNet, backToFirstRank, forcedOrder, time, GPU, g, next[i]));
    }
  } 
  ...
}

到这里就完成了第一次搜索，如前文所述，如果搜索出来的结果没有达到条件，就开始逐步降低条件继续搜索，接下来的过程比较类似，就不再赘述了。

然后看下多机场景下，比如两机十六卡场景，这个时候有网卡，所以ncclTopoSearchParams设置参数为backToFirstRank = -1，backToNet = 7，ncclTopoSearchRec直接执行ncclTopoSearchRecNet。

ncclResult_t ncclTopoSearchRec(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, struct ncclTopoGraph* saveGraph, int* time) {
  int backToNet, backToFirstRank;
  NCCLCHECK(ncclTopoSearchParams(system, graph->pattern, &backToNet, &backToFirstRank));
  if (system->nodes[NET].count) {
    // Start from NET
    ncclTopoSearchRecNet(system, graph, saveGraph, backToNet, backToFirstRank, time);
  }
  ...
}

ncclTopoSearchRecNet会搜索出来一个答案，这里会遍历每个网卡，尝试用每个网卡作为起点搜索环，首先是网卡0，将0写入到inter中第一个channel中，然后将网卡0的带宽减去speedInter，maxChannel减去1，然后后边过程和上述很像，会通过ncclTopoSearchTryGpu搜索出一个环。

ncclResult_t ncclTopoSearchRecNet(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, struct ncclTopoGraph* saveGraph, int backToNet, int backToFirstRank, int* time) {
  const int speed = graph->speedInter;
  for (int n=0; n<system->nodes[NET].count; n++) {
    struct ncclTopoNode* net = system->nodes[NET].nodes+n;
    struct ncclTopoNode* gpu;
    if (graph->collNet && net->net.collSupport == 0) continue;
    if (net->net.width < speed) continue;
    if (net->net.maxChannels == 0) continue;
 
    graph->inter[graph->nChannels*2] = net->id;
    for (int i=0; i<system->nodes[NET].count; i++) {
      if ((system->nodes[NET].nodes[i].net.asic == net->net.asic) &&
          (system->nodes[NET].nodes[i].net.port == net->net.port)) {
        system->nodes[NET].nodes[i].net.width -= speed;
      }
    }
    net->net.maxChannels--;
 
    // First try to replay the last channel
    if (graph->nChannels > 0) {
      int g;
      NCCLCHECK(ncclTopoReplayGetGpu(system, graph, -1, &g));
      NCCLCHECK(ncclTopoSearchTryGpu(system, graph, saveGraph, 0, backToNet, backToFirstRank, FORCED_ORDER_REPLAY, time, NET, n, g));
    }
    if (graph->nChannels == 0 || graph->sameChannels == 0) {
      if (graph->nChannels == 0) {
        // Always try the PCI order first to set a reference, but don't count in the timeout nor let it run for long
        int t = 1 << 10;
        NCCLCHECK(ncclTopoSearchTryGpu(system, graph, saveGraph, 0, backToNet, backToFirstRank, FORCED_ORDER_PCI, &t, NET, n, 0));
        if (t == -1) *time = -1;
      }
      ...
  }
  return ncclSuccess;
}

ncclTopoSearchTryGpu还是会调用ncclTopoSearchRecGpu，当没有遍历完所有GPU节点时，仍然通过递归执行ncclTopoSearchRecGpu来填充graph->intra，最后遍历所有GPU之后step等于7，即backToNet，这里首先拿出来起始网卡，即网卡0，如果搜索参数支持crossNic的话就选一个合法的网卡即可，如果不支持的话就判断网卡0是否合法，合法的话将网卡0填充到graph->inter，一个环就搜索完成了。这里有一个小的疑惑点，在将出口网卡选择好后，并没有将该网卡的带宽减去speed。

ncclResult_t ncclTopoSearchRecGpu(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, struct ncclTopoGraph* saveGraph, struct ncclTopoNode* gpu, int step, int backToNet, int backToFirstRank, int forcedOrder, int *time) {
  if ((*time) <= 0) return ncclSuccess;
  (*time)--;
 
  int ngpus = system->nodes[GPU].count;
  if (step == ngpus) {
    // Determine whether we found a better solution or not
    int copy = 0;
    graph->nChannels++;
    NCCLCHECK(ncclTopoCompareGraphs(graph, saveGraph, ©));
    if (copy) {
      memcpy(saveGraph, graph, sizeof(struct ncclTopoGraph));
      if (graph->nChannels == graph->maxChannels) *time = -1;
    }
    if (graph->nChannels < graph->maxChannels) {
      NCCLCHECK(ncclTopoSearchRec(system, graph, saveGraph, time));
    }
    graph->nChannels--;
    return ncclSuccess;
  }
  graph->intra[graph->nChannels*ngpus+step] = gpu->gpu.rank;
  int g = gpu - system->nodes[GPU].nodes;
  if (step == backToNet) {
    // first get back to NIC
    if (system->nodes[NET].count) {
      int startNetIndex;
      NCCLCHECK(getNetIndex(system, graph->inter[graph->nChannels*2], &startNetIndex));
      struct ncclTopoNode* startNet = system->nodes[NET].nodes+startNetIndex;
      for (int n=0; n<system->nodes[NET].count; n++) {
        struct ncclTopoNode* net = system->nodes[NET].nodes+n;
        if (graph->pattern == NCCL_TOPO_PATTERN_TREE && net->id != startNet->id) continue; // Trees are symmetric
        if (graph->crossNic != 1 && (net->net.asic != startNet->net.asic || net->net.port != startNet->net.port)) continue;
        NCCLCHECK(ncclTopoFollowPath(system, graph, GPU, g, NET, n, 1, &net));
        if (net) {
          graph->inter[graph->nChannels*2+1] = net->id;
          NCCLCHECK(ncclTopoSearchRecGpu(system, graph, saveGraph, gpu, step, -1, backToFirstRank, forcedOrder, time));
          NCCLCHECK(ncclTopoFollowPath(system, graph, GPU, g, NET, n, -1, &net));
        }
      }
    }
  } else if (step < system->nodes[GPU].count-1) {
    // Go to next GPU
    int next[NCCL_TOPO_MAX_NODES];
    int count;
    if (forcedOrder == FORCED_ORDER_PCI) { // Try the PCI order
      next[0] = step+1;
      count = 1;
    } else if (forcedOrder == FORCED_ORDER_REPLAY) { // Try last channel order
      NCCLCHECK(ncclTopoReplayGetGpu(system, graph, step, next));
      count = 1;
    } else { // Normal search
      NCCLCHECK(ncclTopoSearchNextGpuSort(system, graph, gpu, next, &count, backToNet == -1 ? 0 : backToNet == step+1 ? 1 : -1 ));
    }
    for (int i=0; i<count; i++) {
      NCCLCHECK(ncclTopoSearchTryGpu(system, graph, saveGraph, step+1, backToNet, backToFirstRank, forcedOrder, time, GPU, g, next[i]));
    }
  } else if (step == backToFirstRank) {
    ...
  } else {
    // Next path
    NCCLCHECK(ncclTopoSearchRecGpu(system, graph, saveGraph, gpu, ngpus, -1, -1, forcedOrder, time));
  }
  return ncclSuccess;
}

回到ncclTopoSearchRecNet，接下来会尝试复制刚刚搜索出来的环，当搜索出一个答案后，回到第一次ncclTopoSearchRecNet，接下来会尝试从离网卡0最近的GPU开始搜索，而不是从GPU0开始，假设为GPUn，这里会先判断GPUn到PCIe switch的双向带宽是否还有空闲，如果有空闲的话才从GPUn开始搜索。但是和这里的注释表述不太相符，注释的意思是说不会将一个GPU既用来发送，又用来接收（说这种情况会影响带宽，这一点比较疑惑）。

ncclResult_t ncclTopoSearchRecNet(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, struct ncclTopoGraph* saveGraph, int backToNet, int backToFirstRank, int* time) {
  const int speed = graph->speedInter;
  for (int n=0; n<system->nodes[NET].count; n++) {
    ...
      // Then try the most local GPUs
      float maxWidth = 0;
      int minHops = 0xfffffff;
      struct ncclTopoLinkList* paths = net->paths[GPU];
      for (int g=0; g<system->nodes[GPU].count; g++) {
        if (paths[g].width > maxWidth) {
          maxWidth = paths[g].width;
          minHops = paths[g].count;
        } else if (paths[g].width == maxWidth && paths[g].count < minHops) {
          minHops = paths[g].count;
        }
      }
      if (maxWidth >= speed) {
        // In the first loop, avoid using GPUs in both directions between channels (one channel
        // sending from that GPU and one channel receiving to that GPU), since that usually leads
        // to lower BW.
        for (int tryGpuBidir=0; tryGpuBidir<2; tryGpuBidir++) {
          for (int g=0; g<system->nodes[GPU].count; g++) {
            if (paths[g].width == maxWidth && paths[g].count == minHops) {
              gpu = system->nodes[GPU].nodes+g;
              int gpuUsed = gpuPciWidth(gpu) > 0 ? 0 : 1;
              if (tryGpuBidir == gpuUsed) {
                NCCLCHECK(ncclTopoSearchTryGpu(system, graph, saveGraph, 0, backToNet, backToFirstRank, 0, time, NET, n, g));
              }
            }
          }
        }
      }
    }
 
    net->net.maxChannels++;
    for (int i=0; i<system->nodes[NET].count; i++) {
      if ((system->nodes[NET].nodes[i].net.asic == net->net.asic) &&
          (system->nodes[NET].nodes[i].net.port == net->net.port)) {
        system->nodes[NET].nodes[i].net.width += speed;
      }
    }
  }
  return ncclSuccess;
}

到这里就完成了channel的搜索。总结一下，本节就是基于机器拓扑，搜索出一组channel用于数据的通信，并记录到ncclTopoGraph。

（本文经授权后由OneFlow发布。原文：blog.csdn.net/KIDGIN7439/...

欢迎 Star、试用 OneFlow 最新版本：
github.com/Oneflow-Inc...