1# Distributed Data Parallel Benchmark 2 3This tool is used to measure distributed training iteration time. This 4is helpful for evaluating the performance impact of code changes to 5`torch.nn.parallel.DistributedDataParallel`, `torch.distributed`, or 6anything in between. 7 8It optionally produces a JSON file with all measurements, allowing for 9an easy A/B comparison of code, configuration, or environment. This 10comparison can be produced by `diff.py`. 11 12## Requirements 13 14This benchmark depends on PyTorch and torchvision. 15 16## How to run 17 18Run as many copies of this script as you have model replicas. 19 20If you launch a single task per machine with multiple GPUs, consider 21using [`torch.distributed.launch`][launch] to spawn multiple processes 22per machine. 23 24[launch]: https://pytorch.org/docs/stable/distributed.html#launch-utility 25 26Example output (only on rank 0): 27 28``` 29----------------------------------- 30PyTorch distributed benchmark suite 31----------------------------------- 32 33* PyTorch version: 1.4.0a0+05140f0 34* CUDA version: 10.0 35* Distributed backend: nccl 36 37--- nvidia-smi topo -m --- 38 39 GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7 mlx5_2 mlx5_0 mlx5_3 mlx5_1 CPU Affinity 40GPU0 X NV1 NV1 NV2 NV2 SYS SYS SYS SYS PIX SYS PHB 0-19,40-59 41GPU1 NV1 X NV2 NV1 SYS NV2 SYS SYS SYS PIX SYS PHB 0-19,40-59 42GPU2 NV1 NV2 X NV2 SYS SYS NV1 SYS SYS PHB SYS PIX 0-19,40-59 43GPU3 NV2 NV1 NV2 X SYS SYS SYS NV1 SYS PHB SYS PIX 0-19,40-59 44GPU4 NV2 SYS SYS SYS X NV1 NV1 NV2 PIX SYS PHB SYS 0-19,40-59 45GPU5 SYS NV2 SYS SYS NV1 X NV2 NV1 PIX SYS PHB SYS 0-19,40-59 46GPU6 SYS SYS NV1 SYS NV1 NV2 X NV2 PHB SYS PIX SYS 0-19,40-59 47GPU7 SYS SYS SYS NV1 NV2 NV1 NV2 X PHB SYS PIX SYS 0-19,40-59 48mlx5_2 SYS SYS SYS SYS PIX PIX PHB PHB X SYS PHB SYS 49mlx5_0 PIX PIX PHB PHB SYS SYS SYS SYS SYS X SYS PHB 50mlx5_3 SYS SYS SYS SYS PHB PHB PIX PIX PHB SYS X SYS 51mlx5_1 PHB PHB PIX PIX SYS SYS SYS SYS SYS PHB SYS X 52 53Legend: 54 55 X = Self 56 SYS = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI) 57 NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node 58 PHB = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU) 59 PXB = Connection traversing multiple PCIe switches (without traversing the PCIe Host Bridge) 60 PIX = Connection traversing a single PCIe switch 61 NV# = Connection traversing a bonded set of # NVLinks 62 63-------------------------- 64 65 66Benchmark: resnet50 with batch size 32 67 68 sec/iter ex/sec sec/iter ex/sec sec/iter ex/sec sec/iter ex/sec 69 1 GPUs -- no ddp: p50: 0.097s 329/s p75: 0.097s 329/s p90: 0.097s 329/s p95: 0.097s 329/s 70 1 GPUs -- 1M/1G: p50: 0.100s 319/s p75: 0.100s 318/s p90: 0.100s 318/s p95: 0.100s 318/s 71 2 GPUs -- 1M/2G: p50: 0.103s 310/s p75: 0.103s 310/s p90: 0.103s 310/s p95: 0.103s 309/s 72 4 GPUs -- 1M/4G: p50: 0.103s 310/s p75: 0.103s 310/s p90: 0.103s 310/s p95: 0.103s 310/s 73 8 GPUs -- 1M/8G: p50: 0.104s 307/s p75: 0.104s 307/s p90: 0.104s 306/s p95: 0.104s 306/s 74 16 GPUs -- 2M/8G: p50: 0.104s 306/s p75: 0.104s 306/s p90: 0.104s 306/s p95: 0.104s 306/s 75 76Benchmark: resnet101 with batch size 32 77 78 sec/iter ex/sec sec/iter ex/sec sec/iter ex/sec sec/iter ex/sec 79 1 GPUs -- no ddp: p50: 0.162s 197/s p75: 0.162s 197/s p90: 0.162s 197/s p95: 0.162s 197/s 80 1 GPUs -- 1M/1G: p50: 0.171s 187/s p75: 0.171s 186/s p90: 0.171s 186/s p95: 0.172s 185/s 81 2 GPUs -- 1M/2G: p50: 0.176s 182/s p75: 0.176s 181/s p90: 0.176s 181/s p95: 0.176s 181/s 82 4 GPUs -- 1M/4G: p50: 0.176s 182/s p75: 0.176s 181/s p90: 0.176s 181/s p95: 0.176s 181/s 83 8 GPUs -- 1M/8G: p50: 0.179s 179/s p75: 0.179s 178/s p90: 0.180s 178/s p95: 0.180s 177/s 84 16 GPUs -- 2M/8G: p50: 0.179s 178/s p75: 0.180s 177/s p90: 0.183s 174/s p95: 0.188s 170/s 85 86Benchmark: resnext50_32x4d with batch size 32 87 88 sec/iter ex/sec sec/iter ex/sec sec/iter ex/sec sec/iter ex/sec 89 1 GPUs -- no ddp: p50: 0.145s 220/s p75: 0.145s 220/s p90: 0.145s 220/s p95: 0.145s 220/s 90 1 GPUs -- 1M/1G: p50: 0.147s 217/s p75: 0.147s 217/s p90: 0.148s 216/s p95: 0.148s 216/s 91 2 GPUs -- 1M/2G: p50: 0.153s 209/s p75: 0.153s 209/s p90: 0.153s 209/s p95: 0.153s 209/s 92 4 GPUs -- 1M/4G: p50: 0.153s 208/s p75: 0.153s 208/s p90: 0.154s 208/s p95: 0.154s 208/s 93 8 GPUs -- 1M/8G: p50: 0.157s 204/s p75: 0.157s 204/s p90: 0.157s 203/s p95: 0.157s 203/s 94 16 GPUs -- 2M/8G: p50: 0.157s 203/s p75: 0.157s 203/s p90: 0.158s 203/s p95: 0.158s 202/s 95 96Benchmark: resnext101_32x8d with batch size 32 97 98 sec/iter ex/sec sec/iter ex/sec sec/iter ex/sec sec/iter ex/sec 99 1 GPUs -- no ddp: p50: 0.415s 77/s p75: 0.415s 77/s p90: 0.416s 76/s p95: 0.417s 76/s 100 1 GPUs -- 1M/1G: p50: 0.425s 75/s p75: 0.426s 75/s p90: 0.426s 75/s p95: 0.426s 75/s 101 2 GPUs -- 1M/2G: p50: 0.438s 73/s p75: 0.439s 72/s p90: 0.439s 72/s p95: 0.439s 72/s 102 4 GPUs -- 1M/4G: p50: 0.439s 72/s p75: 0.439s 72/s p90: 0.440s 72/s p95: 0.440s 72/s 103 8 GPUs -- 1M/8G: p50: 0.447s 71/s p75: 0.447s 71/s p90: 0.448s 71/s p95: 0.448s 71/s 104 16 GPUs -- 2M/8G: p50: 0.450s 71/s p75: 0.451s 70/s p90: 0.451s 70/s p95: 0.451s 70/s 105``` 106 107## How to diff 108 109Run the benchmark with the `--json PATH_TO_REPORT_FILE` argument to 110produce the JSON file that the diff script can consume. 111 112Then, run the diff script as follows: 113 114``` 115$ python3 diff.py PATH_TO_BASELINE_FILE PATH_TO_TEST_FILE 116 baseline test 117 -------------------- -------------------- 118bucket_size: 25 vs 1 119cuda_version: 10.0 vs 10.0 120distributed_backend: nccl vs nccl 121pytorch_version: 1.4.0a0+05140f0 vs 1.4.0a0+05140f0 122 123Benchmark: resnet50 with batch size 32 124 125 sec/iter ex/sec diff sec/iter ex/sec diff 126 1 GPUs: p75: 0.101s 317/s -0.3% p95: 0.101s 317/s -0.4% 127 2 GPUs: p75: 0.104s 306/s -1.0% p95: 0.104s 306/s -1.0% 128 4 GPUs: p75: 0.105s 305/s -1.6% p95: 0.105s 304/s -1.8% 129 8 GPUs: p75: 0.107s 299/s -2.6% p95: 0.107s 298/s -2.7% 130 16 GPUs: p75: 0.108s 294/s -3.8% p95: 0.122s 262/s -16.4% 131 132Benchmark: resnet101 with batch size 32 133 134 sec/iter ex/sec diff sec/iter ex/sec diff 135 1 GPUs: p75: 0.172s 185/s -1.2% p95: 0.172s 185/s -1.3% 136 2 GPUs: p75: 0.179s 178/s -2.1% p95: 0.179s 178/s -2.0% 137 4 GPUs: p75: 0.180s 177/s -2.6% p95: 0.180s 177/s -2.6% 138 8 GPUs: p75: 0.184s 173/s -3.5% p95: 0.184s 173/s -3.5% 139 16 GPUs: p75: 0.187s 170/s -0.1% p95: 0.204s 157/s -7.9% 140 141Benchmark: resnext50_32x4d with batch size 32 142 143 sec/iter ex/sec diff sec/iter ex/sec diff 144 1 GPUs: p75: 0.149s 214/s -1.0% p95: 0.149s 214/s -0.9% 145 2 GPUs: p75: 0.156s 205/s -1.5% p95: 0.156s 205/s -1.6% 146 4 GPUs: p75: 0.156s 204/s -1.6% p95: 0.157s 204/s -1.8% 147 8 GPUs: p75: 0.159s 200/s -1.5% p95: 0.159s 200/s -1.5% 148 16 GPUs: p75: 0.161s 198/s -1.9% p95: 0.162s 197/s -2.3% 149 150Benchmark: resnext101_32x8d with batch size 32 151 152 sec/iter ex/sec diff sec/iter ex/sec diff 153 1 GPUs: p75: 0.427s 74/s -0.8% p95: 0.428s 74/s -0.7% 154 2 GPUs: p75: 0.444s 72/s -1.3% p95: 0.445s 71/s -0.7% 155 4 GPUs: p75: 0.444s 72/s -1.1% p95: 0.445s 71/s -0.8% 156 8 GPUs: p75: 0.452s 70/s -1.3% p95: 0.452s 70/s -1.3% 157 16 GPUs: p75: 0.455s 70/s -0.7% p95: 0.456s 70/s -0.6% 158``` 159 160This compares throughput between `bucket_cap_mb=25` (the default) and 161`bucket_cap_mb=1` on 8 DGX machines with V100 GPUs. It confirms that 162even for a relatively small model on machines with a very fast 163interconnect (4x 100Gb InfiniBand per machine), it still pays off to 164batch allreduce calls. 165