ps/grpc/grpc_ps_bench.cpp
| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include <gflags/gflags.h> | ||
| 2 | |||
| 3 | #include <algorithm> | ||
| 4 | #include <chrono> | ||
| 5 | #include <cstdio> | ||
| 6 | #include <vector> | ||
| 7 | |||
| 8 | #include "grpc_ps_client.h" | ||
| 9 | |||
| 10 | DEFINE_string(host, "127.0.0.1", "PS server host"); | ||
| 11 | DEFINE_int32(port, 15000, "PS server port"); | ||
| 12 | DEFINE_int32(warmup, 10, "Warmup iterations"); | ||
| 13 | DEFINE_int32(iters, 100, "Measured iterations per payload size"); | ||
| 14 | |||
| 15 | static const int64_t kSizes[] = { | ||
| 16 | 4 * 1024, // 4 KB | ||
| 17 | 64 * 1024, // 64 KB | ||
| 18 | 512 * 1024, // 512 KB | ||
| 19 | 4 * 1024 * 1024, // 4 MB | ||
| 20 | }; | ||
| 21 | |||
| 22 | ✗ | static double now_us() { | |
| 23 | ✗ | return std::chrono::duration_cast<std::chrono::nanoseconds>( | |
| 24 | ✗ | std::chrono::steady_clock::now().time_since_epoch()) | |
| 25 | ✗ | .count() / | |
| 26 | ✗ | 1e3; | |
| 27 | } | ||
| 28 | |||
| 29 | template <typename Fn> | ||
| 30 | static void | ||
| 31 | ✗ | RunBench(const char* label, int64_t size_bytes, int warmup, int iters, Fn fn) { | |
| 32 | ✗ | for (int i = 0; i < warmup; i++) | |
| 33 | ✗ | fn(size_bytes); | |
| 34 | |||
| 35 | ✗ | std::vector<double> latencies(iters); | |
| 36 | ✗ | double t_start = now_us(); | |
| 37 | ✗ | for (int i = 0; i < iters; i++) { | |
| 38 | ✗ | double t0 = now_us(); | |
| 39 | ✗ | fn(size_bytes); | |
| 40 | ✗ | latencies[i] = now_us() - t0; | |
| 41 | } | ||
| 42 | ✗ | double elapsed_us = now_us() - t_start; | |
| 43 | |||
| 44 | ✗ | std::sort(latencies.begin(), latencies.end()); | |
| 45 | ✗ | double avg_us = elapsed_us / iters; | |
| 46 | ✗ | double p50_us = latencies[iters * 50 / 100]; | |
| 47 | ✗ | double p99_us = latencies[iters * 99 / 100]; | |
| 48 | ✗ | double tput_mb = | |
| 49 | ✗ | (static_cast<double>(size_bytes) * iters / (1024.0 * 1024.0)) / | |
| 50 | ✗ | (elapsed_us / 1e6); | |
| 51 | |||
| 52 | ✗ | printf(" [%-4s] %6.0f KB avg=%8.1f us p50=%8.1f us p99=%8.1f us " | |
| 53 | "tput=%8.1f MB/s\n", | ||
| 54 | label, | ||
| 55 | size_bytes / 1024.0, | ||
| 56 | avg_us, | ||
| 57 | p50_us, | ||
| 58 | p99_us, | ||
| 59 | tput_mb); | ||
| 60 | ✗ | } | |
| 61 | |||
| 62 | ✗ | int main(int argc, char** argv) { | |
| 63 | ✗ | gflags::ParseCommandLineFlags(&argc, &argv, true); | |
| 64 | |||
| 65 | ✗ | GRPCParameterClient client(FLAGS_host, FLAGS_port, /*shard=*/0); | |
| 66 | |||
| 67 | ✗ | printf("PS Network Benchmark server=%s:%d warmup=%d iters=%d\n\n", | |
| 68 | FLAGS_host.c_str(), | ||
| 69 | FLAGS_port, | ||
| 70 | FLAGS_warmup, | ||
| 71 | FLAGS_iters); | ||
| 72 | |||
| 73 | ✗ | printf("LoadFakeData (download: server -> client)\n"); | |
| 74 | ✗ | for (int64_t sz : kSizes) | |
| 75 | ✗ | RunBench("load", sz, FLAGS_warmup, FLAGS_iters, [&](int64_t n) { | |
| 76 | ✗ | client.LoadFakeData(n); | |
| 77 | ✗ | }); | |
| 78 | |||
| 79 | ✗ | printf("\nDumpFakeData (upload: client -> server)\n"); | |
| 80 | ✗ | for (int64_t sz : kSizes) | |
| 81 | ✗ | RunBench("dump", sz, FLAGS_warmup, FLAGS_iters, [&](int64_t n) { | |
| 82 | ✗ | client.DumpFakeData(n); | |
| 83 | ✗ | }); | |
| 84 | |||
| 85 | ✗ | return 0; | |
| 86 | ✗ | } | |
| 87 |