// one second of wall time), it should remove "a" and leave "b" alone.
  uint64 start = Env::Default()->NowSeconds();
  do {
    Env::Default()->SleepForMicroseconds(100000);
  } while (cache.CacheSize() == 24 && Env::Default()->NowSeconds() - start < 3);
  // There should be one block left in the cache, and it should be the first
  // block of "b".
  EXPECT_EQ(cache.CacheSize(), 8);

  absl::MutexLock lock(&mu_);
  return cache_size_;
}

void RamFileBlockCache::Prune() {
  while (!stop_pruning_thread_.WaitForNotificationWithTimeout(
      absl::Microseconds(1000000))) {
    absl::MutexLock lock(&mu_);
    uint64_t now = timer_seconds_();
    while (!lra_list_.empty()) {
      auto it = block_map_.find(lra_list_.back());
      if (now - it->second->timestamp <= max_staleness_) {

                  server_def_0, tensorflow::Env::Default(), &worker_server2)
                  .ok());
  ASSERT_TRUE(worker_server2->Start().ok());

  // Create two contexts.
  int32_t init_timeout_in_ms = 300000;
  TFE_Context* ctx_0 =
      CreateContext(serialized_server_def_0,
                    /*isolate_session_state=*/false, init_timeout_in_ms);
  TFE_Context* ctx_1 =
      CreateContext(serialized_server_def_1,

    // This is usually unnecessary, but we've experienced the occasional test
    // failure when testing async mode with no explicit forwarding.
    for (int i = 0; i < 100000; ++i) {
      TFE_Op* add_op_dummy = AddOp(ctx, m, m);
      TFE_OpSetDevice(add_op_dummy, cpu_device_name.c_str(), status);
      ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);