Tencent/flare

关于调度参数的疑问

yinghaoyu opened this issue · 1 comments

yinghaoyu commented

请问一下，这里针对不同策略的计算是有什么依据吗（根据什么原理确定的参数）？

SchedulingParameters GetSchedulingParameters(SchedulingProfile profile,
                                             std::size_t numa_domains,
                                             std::size_t available_processors,
                                             std::size_t desired_concurrency) {
  if (profile == SchedulingProfile::ComputeHeavy) {
    return GetSchedulingParametersForComputeHeavy(desired_concurrency);
  } else if (profile == SchedulingProfile::Compute) {
    return GetSchedulingParametersForCompute(numa_domains, available_processors,
                                             desired_concurrency);
  } else if (profile == SchedulingProfile::Neutral) {
    // @sa: `SchedulingProfile` for the constants below.
    return GetSchedulingParametersOfGroupSize(numa_domains, desired_concurrency,
                                              16, 32);
  } else if (profile == SchedulingProfile::Io) {
    return GetSchedulingParametersOfGroupSize(numa_domains, desired_concurrency,
                                              12, 24);
  } else if (profile == SchedulingProfile::IoHeavy) {
    return GetSchedulingParametersOfGroupSize(numa_domains, desired_concurrency,
                                              8, 16);
  }
  FLARE_UNREACHABLE("Unexpected scheduling profile [{}].",
                    underlying_value(profile));
}
0x804d8000 commented

计算密集的一般有一定的处理耗时，为了避免worker不均衡导致的长尾请求，优先考虑更大的group size。io密集的因为都处理的很快，可以容忍一定的不均衡，所以优先考虑调度效率。

  • 👍1