NCCL源码解析④：建图过程

作者｜KIDGINBROOK
更新｜潘丽晨

上次分析了NCCL对机器PCI系统进行拓扑分析的过程，产出的结果为xml格式，接下来，NCCL会根据这个xml进图的建立过程以便之后进行路径搜索。

ncclTopoGetSystem的最后会执行ncclTopoGetSystemFromXml将xml格式转成图格式。

ncclResult_t ncclTopoGetSystemFromXml(struct ncclXml* xml, struct ncclTopoSystem** topoSystem) {
  NCCLCHECK(ncclCalloc(topoSystem, 1));
  struct ncclXmlNode* topNode;
  NCCLCHECK(xmlFindTag(xml, "system", &topNode));
  for (int s=0; snSubs; s++) {
    struct ncclXmlNode* node = topNode->subs[s];
    if (strcmp(node->name, "cpu") == 0) NCCLCHECK(ncclTopoAddCpu(node, *topoSystem));
  }
  NCCLCHECK(ncclTopoAddNvLinks(topNode, *topoSystem, NULL));
 
  NCCLCHECK(ncclTopoConnectCpus(*topoSystem));
  NCCLCHECK(ncclTopoSortSystem(*topoSystem));
 
  return ncclSuccess;
}

从xml中拿到根节点"system"，然后遍历子节点中的"cpu"，对每个cpu通过ncclTopoAddCpu进行建图，这里一个cpu其实就是一个numa。

ncclResult_t ncclTopoAddCpu(struct ncclXmlNode* xmlCpu, struct ncclTopoSystem* system) {
  int numaId;
  NCCLCHECK(xmlGetAttrInt(xmlCpu, "numaid", &numaId));
  struct ncclTopoNode* cpu;
  NCCLCHECK(ncclTopoCreateNode(system, &cpu, CPU, numaId));
  const char* str;
  NCCLCHECK(xmlGetAttr(xmlCpu, "affinity", &str));
  if (str != NULL) {
    NCCLCHECK(ncclStrToCpuset(str, &cpu->cpu.affinity));
  }
 
  NCCLCHECK(xmlGetAttrStr(xmlCpu, "arch", &str));
  NCCLCHECK(kvConvertToInt(str, &cpu->cpu.arch, kvDictCpuArch));
  if (cpu->cpu.arch == NCCL_TOPO_CPU_ARCH_X86) {
    NCCLCHECK(xmlGetAttrStr(xmlCpu, "vendor", &str));
    NCCLCHECK(kvConvertToInt(str, &cpu->cpu.vendor, kvDictCpuVendor));
    if (cpu->cpu.vendor == NCCL_TOPO_CPU_VENDOR_INTEL) {
      int familyId, modelId;
      NCCLCHECK(xmlGetAttrInt(xmlCpu, "familyid", &familyId));
      NCCLCHECK(xmlGetAttrInt(xmlCpu, "modelid", &modelId));
      cpu->cpu.model = (familyId == 6 && modelId >= 0x55) ? NCCL_TOPO_CPU_TYPE_SKL : NCCL_TOPO_CPU_INTEL_BDW;
    }   
  }
  for (int s=0; snSubs; s++) {
    struct ncclXmlNode* node = xmlCpu->subs[s];
    if (strcmp(node->name, "pci") == 0) NCCLCHECK(ncclTopoAddPci(node, system, cpu));
    if (strcmp(node->name, "nic") == 0) {
      struct ncclTopoNode* nic = NULL;
      NCCLCHECK(ncclTopoGetNode(system, &nic, NIC, 0));
      if (nic == NULL) {
        NCCLCHECK(ncclTopoCreateNode(system, &nic, NIC, 0));
        NCCLCHECK(ncclTopoConnectNodes(cpu, nic, LINK_PCI, LOC_WIDTH));
        NCCLCHECK(ncclTopoConnectNodes(nic, cpu, LINK_PCI, LOC_WIDTH));
      }   
      NCCLCHECK(ncclTopoAddNic(node, system, nic));
    }   
  }
  return ncclSuccess;
}

接着创建一个cpu node，id为numaid，设置cpu的affinity，即该numa对应的核，设置cpu对应vendor等信息。

然后遍历cpu node的子节点，根据不同的类型执行不同的函数，如果是PCI节点，则执行ncclTopoAddPci。

ncclResult_t ncclTopoAddPci(struct ncclXmlNode* xmlPci, struct ncclTopoSystem* system, struct ncclTopoNode* parent) {
  const char* str;
 
  int type;
  NCCLCHECK(xmlGetAttrStr(xmlPci, "class", &str));
  NCCLCHECK(kvConvertToInt(str, &type, kvDictPciClass));
 
  int64_t busId;
  NCCLCHECK(xmlGetAttrStr(xmlPci, "busid", &str));
  NCCLCHECK(busIdToInt64(str, &busId));
 
  struct ncclTopoNode* node = NULL;
  if (type == GPU) {
    struct ncclXmlNode* xmlGpu;
    NCCLCHECK(xmlGetSub(xmlPci, "gpu", &xmlGpu));
    if (xmlGpu == NULL) return ncclSuccess;
    int index;
    NCCLCHECK(xmlGetAttrIndex(xmlGpu, "rank", &index));
    if (index == -1) return ncclSuccess;
    NCCLCHECK(ncclTopoCreateNode(system, &node, type, busId));
    NCCLCHECK(ncclTopoAddGpu(xmlGpu, system, node));
  }
  if (type == NIC) {
    struct ncclXmlNode* xmlNic;
    NCCLCHECK(xmlGetSub(xmlPci, "nic", &xmlNic));
    if (xmlNic == NULL) return ncclSuccess;
 
    // Ignore sub device ID and merge multi-port NICs into one PCI device.
    busId &= 0xfffffffffffffff0;
    struct ncclTopoNode* nicNode = NULL;
    NCCLCHECK(ncclTopoGetNode(system, &nicNode, type, busId));
    if (nicNode == NULL) {
      NCCLCHECK(ncclTopoCreateNode(system, &nicNode, type, busId));
      node = nicNode; // Connect it to parent later on
    }
    NCCLCHECK(ncclTopoAddNic(xmlNic, system, nicNode));
  } else if (type == PCI) {
    NCCLCHECK(ncclTopoCreateNode(system, &node, type, busId));
    for (int s=0; snSubs; s++) {
      struct ncclXmlNode* xmlSubPci = xmlPci->subs[s];
      NCCLCHECK(ncclTopoAddPci(xmlSubPci, system, node));
    }
  }
 
  if (node) {
    int width, speed;
    NCCLCHECK(xmlGetAttrInt(xmlPci, "link_width", &width));
    NCCLCHECK(xmlGetAttrStr(xmlPci, "link_speed", &str));
 
    // Manage cases where speed was not indicated in /sys
    if (width == 0) width = 16;
    NCCLCHECK(kvConvertToInt(str, &speed, kvDictPciGen)); // Values in 100Mbps, per lane (we want GB/s in the end)
 
    NCCLCHECK(ncclTopoConnectNodes(node, parent, LINK_PCI, width*speed/80.0));
    NCCLCHECK(ncclTopoConnectNodes(parent, node, LINK_PCI, width*speed/80.0));
  }
  return ncclSuccess;
}

首先获取pci的type和busId， 然后判断type，如果是PCI，那么创建一个PCI node，递归执行ncclTopoAddPci，直到遇到NIC或者GPU xml节点。

如果遇到的是NIC，那么创建NIC节点，然后执行ncclTopoAddNic，这里会在xml nic下遍历xml net，对每个xml net创建net node，id为dev，然后设置speed，port，gdr等属性。

ncclResult_t ncclTopoAddNet(struct ncclXmlNode* xmlNet, struct ncclTopoSystem* system, struct ncclTopoNode* nic) {
  int dev;
  NCCLCHECK(xmlGetAttrInt(xmlNet, "dev", &dev));
 
  struct ncclTopoNode* net;
  NCCLCHECK(ncclTopoCreateNode(system, &net, NET, dev));
  const char* str;
  NCCLCHECK(xmlGetAttr(xmlNet, "guid", &str));
  if (str) sscanf(str, "0x%lx", &net->net.asic);
  else net->net.asic = dev;
 
  ncclDebugNoWarn = NCCL_GRAPH;
  int mbps;
  if (xmlGetAttrInt(xmlNet, "speed", &mbps) != ncclSuccess) mbps = 0;
  if (mbps <= 0) mbps = 10000; // Some NICs define speed = -1
  net->net.width = mbps / 8000.0;
  if (xmlGetAttrInt(xmlNet, "port", &net->net.port) != ncclSuccess) net->net.port = 0;
  if (xmlGetAttrInt(xmlNet, "gdr", &net->net.gdrSupport) != ncclSuccess) net->net.gdrSupport = 0;
  if (xmlGetAttrInt(xmlNet, "maxconn", &net->net.maxChannels) != ncclSuccess) net->net.maxChannels = MAXCHANNELS;
  if (xmlGetAttrInt(xmlNet, "coll", &net->net.collSupport) != ncclSuccess) net->net.collSupport = 0;
  ncclDebugNoWarn = 0;
 
  NCCLCHECK(ncclTopoConnectNodes(nic, net, LINK_NET, net->net.width));
  NCCLCHECK(ncclTopoConnectNodes(net, nic, LINK_NET, net->net.width));
  return ncclSuccess;
}
 
ncclResult_t ncclTopoAddNic(struct ncclXmlNode* xmlNic, struct ncclTopoSystem* system, struct ncclTopoNode* nic) {
  for (int s=0; snSubs; s++) {
    struct ncclXmlNode* xmlNet = xmlNic->subs[s];
    if (strcmp(xmlNet->name, "net") != 0) continue;
    int index;
    NCCLCHECK(xmlGetAttrIndex(xmlNet, "dev", &index));
    if (index == -1) continue;
    NCCLCHECK(ncclTopoAddNet(xmlNet, system, nic));
  }
  return ncclSuccess;
}

然后通过建立net node到nic node的正反向边，设置边的类型，边上累计带宽，并且当前节点的边按照带宽从大到小排序。

ncclResult_t ncclTopoConnectNodes(struct ncclTopoNode* node, struct ncclTopoNode* remNode, int type, float width) {
  // Aggregate links into higher width for NVLink
  struct ncclTopoLink* link;
  for (link = node->links; link->remNode; link++) {
    if (link->remNode == remNode && link->type == type) break;
  }
  if (link->remNode == NULL) node->nlinks++;
  link->type = type;
  link->remNode = remNode;
  link->width += width;
 
  // Sort links in BW descending order
  struct ncclTopoLink linkSave;
  memcpy(&linkSave, link, sizeof(struct ncclTopoLink));
  while (link != node->links) {
    if ((link-1)->width >= linkSave.width) break;
    memcpy(link, link-1, sizeof(struct ncclTopoLink));
    link--;
  }
  memcpy(link, &linkSave, sizeof(struct ncclTopoLink));
  return ncclSuccess;
}

到这里就添加完成了NIC，回到ncclTopoAddPci里，如果是gpu的话则创建gpu node，然后设置gpu node的rank，dev，gdr等属性。最后通过ncclTopoConnectNodes建立当前节点到子节点的双向边。

到这里就完成了每个numa节点下的建图，然后开始添加nvlink和QPI以连接，先看下nvlink。

ncclResult_t ncclTopoAddNvLinks(struct ncclXmlNode* node, struct ncclTopoSystem* system, const char* parentBusId) {
  if (strcmp(node->name, "nvlink") == 0) {
    struct ncclTopoNode* gpu = NULL;
    int64_t pBusId;
    NCCLCHECK(busIdToInt64(parentBusId, &pBusId));
    NCCLCHECK(ncclTopoGetNode(system, &gpu, GPU, pBusId));
    if (gpu == NULL) {
      WARN("Add NVLink error : could not find GPU %lx\n", pBusId);
      return ncclInternalError;
    }
    int count;
    NCCLCHECK(xmlGetAttrInt(node, "count", &count));
    const char* targetClass;
    NCCLCHECK(xmlGetAttrStr(node, "tclass", &targetClass));
    int targetType;
    NCCLCHECK(kvConvertToInt(targetClass, &targetType, kvDictPciClass));
    struct ncclTopoNode* remote = NULL;
    if (targetType == GPU) {
      // NVL P2P connection to another GPU
      const char* target;
      NCCLCHECK(xmlGetAttrStr(node, "target", &target));
      int64_t busId;
      NCCLCHECK(busIdToInt64(target, &busId));
      NCCLCHECK(ncclTopoGetNode(system, &remote, GPU, busId));
    } else if (targetType == CPU) {
      // NVL connection to the local CPU
      NCCLCHECK(findLocalCpu(gpu, &remote));
    } else {
      if (system->nodes[NVS].count == 0) {
        NCCLCHECK(ncclTopoCreateNode(system, &remote, NVS, 0));
      } else {
        remote = system->nodes[NVS].nodes;
      }
    }
    if (remote) {
      int nvlSpeed = gpu->gpu.cudaCompCap == 60 ? PASCAL_NVLINK_WIDTH : VOLTA_NVLINK_WIDTH;
      NCCLCHECK(ncclTopoConnectNodes(gpu, remote, LINK_NVL, count*nvlSpeed));
      if (remote->type != GPU) {
        NCCLCHECK(ncclTopoConnectNodes(remote, gpu, LINK_NVL, count*nvlSpeed));
      }
    }
  } else {
    const char* busId;
    NCCLCHECK(xmlGetAttr(node, "busid", &busId));
    for (int s=0; snSubs; s++) {
      NCCLCHECK(ncclTopoAddNvLinks(node->subs[s], system, busId ? busId : parentBusId));
    }
  }
  return ncclSuccess;
}

从根节点递归遍历下去，直到遇到nvlink xml节点，然后拿到nvlink的父节点，即gpu节点，然后通过tclass获取对端PCI设备类型，如果是gpu或者cpu，直接返回对端node，如果是nvswitch，那就先创建nvswitch节点，然后创建当前gpu节点和对端的双向边。然后通过ncclTopoConnectCpus将cpu两两连接。

最后为了方便后续搜索channel，通过ncclTopoSort递归将每个PCI节点的边按照nvlink，向下的PCI连接，向上的PCI连接，QPI的顺序进行排序，因为建边的过程中已经按照带宽排序过，所以nvlink一定在最前边，QPI一定在最后，因此只需要对中间的PCI排序即可。

static ncclResult_t ncclTopoSort(struct ncclTopoNode* node, struct ncclTopoNode* upNode) {
  // Shift all links to have upLink as last link
  if (upNode) {
    int l=0;
    while (node->links[l].remNode != upNode) l++;
    struct ncclTopoLink upLink;
    memcpy(&upLink, node->links+l, sizeof(struct ncclTopoLink));
    while (node->links[l+1].remNode) {
      memcpy(node->links+l, node->links+l+1, sizeof(struct ncclTopoLink));
      l++;
    }   
    memcpy(node->links+l, &upLink, sizeof(struct ncclTopoLink));
  }
 
  // Recursively sort the PCI tree
  for (int l=0; lnlinks; l++) {
    struct ncclTopoLink* link = node->links+l;
    if (link->type == LINK_PCI && link->remNode != upNode) NCCLCHECK(ncclTopoSort(link->remNode, node));
  }
  return ncclSuccess;
}

到这里就完成了整个的建图过程。总结下，由于拓扑分析产出的xml不便于进行后续的路径搜索，所以本节基于xml对PCI系统进行了建图。

其他人都在看

• 关于语言大模型的八大论断
  

• 揭示GPT Tokenizer的工作原理

• 语言大模型100K上下文窗口的秘诀
  

• GPT总设计师：大型语言模型的未来

• NCCL源码解析③：机器内拓扑分析
  

• OneEmbedding:单卡训练TB级推荐模型不是梦

• GLM训练加速：性能最高提升3倍，显存节省1/3

试用OneFlow: github.com/Oneflow-Inc/oneflow/