RefreshRemoteAttributes() is used to initialize device attributes between workers when we use CollectiveOp. We need it to send GetStatus RPC with fail_fast = false so that each worker can block waiting for other workers to start up.

PiperOrigin-RevId: 266207854
2019-08-29 12:57:42 -07:00 · 2019-08-29 12:57:42 -07:00 · 42652b1699
commit 42652b1699
parent 988882bb84
1 changed files with 1 additions and 1 deletions
--- a/tensorflow/core/distributed_runtime/device_resolver_distributed.cc
+++ b/tensorflow/core/distributed_runtime/device_resolver_distributed.cc
@ -98,7 +98,7 @@ void DeviceResolverDistributed::RefreshRemoteAttributes(
  WorkerInterface* worker = worker_cache_->GetOrCreateWorker(task);
  CHECK(worker) << "Failed to get worker for " << task;
  worker->GetStatusAsync(
-      req, resp, /*fail_fast=*/true,
+      req, resp, /*fail_fast=*/false,
      [this, device, task, req, resp, worker, done](Status s) {
        if (s.ok()) {
          mutex_lock l(mu_);