Is there a relationship between ncclTopoSearchNextGpuSort and followPath

[zhangdexin]

In search.c:
I found one main effect in followPath func is the path->list[step].bw subtract fwBw:

static ncclResult_t followPath(struct ncclTopoLinkList* path, struct ncclTopoNode* start, int maxSteps, float bw, int* steps) {
  // ...

  struct ncclTopoNode* node = start;
  for (int step=0; step<maxSteps; step++) {
    struct ncclTopoLink* link = path->list[step];
    struct ncclTopoLink* revLink = NULL;
    float fwBw = link->type == LINK_PCI ? pciBw : bw;

    // ...
    
    SUB_ROUND(link->bw, fwBw);                                // <<  here
    node = link->remNode;
  }
  *steps = maxSteps;
  return ncclSuccess;
}

And I found ncclTopoSearchNextGpuSort func that use path.bw(not path.link.bw) cal score?
In that case, each calculation is the same and will not be dynamically adjusted based on bandwidth consumption?
Is there a relationship between ncclTopoSearchNextGpuSort and followPath？
Do you need to add path.bw subtract fwBw in followPath?

---

followPath 函数中减去的是path.link.bw
然后我在ncclTopoSearchNextGpuSort 函数中发现计算分数使用的path.bw，那这样每次迭代都是一样的，是不是有问题呢？
是不是需要在followPath 函数增加一个减少path.bw的代码？