本文為英文版的機器翻譯版本，如內容有任何歧義或不一致之處，概以英文版為準。 # 在 DLAMI 上使用 EFA 下一節說明如何使用 EFA 在 上執行多節點應用程式 AWS 深度學習 AMIs。 ## 使用 EFA 執行多節點應用程式 若要跨節點叢集執行應用程式，需要下列組態 **Topics** + [ ### 啟用無密碼 SSH ](#tutorial-efa-using-multi-node-ssh) + [ ### 建立主機檔案 ](#tutorial-efa-using-multi-node-hosts) + [ ### NCCL 測試 ](#tutorial-efa-using-2node) ### 啟用無密碼 SSH 在叢集中選取一個節點做為領導節點。其餘的節點稱為成員節點。 1. 在領導節點上，產生 RSA 金鑰對。 ``` ssh-keygen -t rsa -N "" -f ~/.ssh/id_rsa ``` 1. 變更領導節點上私有金鑰的許可。 ``` chmod 600 ~/.ssh/id_rsa ``` 1. 將公有金鑰複製到 ，`~/.ssh/id_rsa.pub`並將其附加到叢集中成員節點`~/.ssh/authorized_keys`的 。 1. 您現在應該可以使用私有 IP 直接從領導節點登入到成員節點。 ``` ssh ``` 1. 在領導節點上的 \$1/.ssh/config 檔案中加入以下內容，以停用 strictHostKeyChecking 並啟用領導節點上的代理程式轉送： ``` Host * ForwardAgent yes Host * StrictHostKeyChecking no ``` 1. 在 Amazon Linux 2 執行個體上，在領導節點上執行下列命令，以提供組態檔案的正確許可： ``` chmod 600 ~/.ssh/config ``` ### 建立主機檔案 在領導節點上，建立主機檔案以識別叢集中的節點。主機檔案對於叢集中每個節點都必須有項目。建立一個檔案 \$1/hosts，並使用私有 IP 新增每個節點，如下所示： ``` localhost slots=8 slots=8 slots=8 ``` ### NCCL 測試 **注意** 這些測試已使用 EFA 1.38.0 版和 OFI NCCL 外掛程式 1.13.2 執行。 下列是 Nvidia 提供的 NCCL 測試子集，用於測試多個運算節點的功能和效能 **支援的執行個體：P3dn, P4, P5, P5e, P5en** #### 效能測試 ##### P4d.24xlarge 上的多節點 NCCL 效能測試 若要使用 EFA 檢查 NCCL 效能，請執行官方 NCCL[-Tests 儲存庫上可用的標準 NCCL](https://github.com/NVIDIA/nccl-tests.git) 效能測試。DLAMI 隨附針對 CUDA XX.X 建置的測試。 您可以同樣使用 EFA 執行自己的指令碼。 建構您自己的指令碼時，請參閱下列指引： + 使用 EFA 執行 NCCL 應用程式時，請使用 mpirun 的完整路徑，如範例所示。 + 根據叢集中的執行個體和 GPU 的數目來變更參數 np 和 N。 + 新增 NCCL\$1DEBUG=INFO 旗標，並確保日誌指出 EFA 用量為 "Selected Provider is EFA"。 + 將訓練日誌位置設定為剖析以進行驗證 ``` TRAINING_LOG="testEFA_$(date +"%N").log" ``` 在任何成員節點上使用命令 `watch nvidia-smi` 來監視 GPU 使用量。下列`watch nvidia-smi`命令適用於一般 CUDA xx.x 版本，並取決於執行個體的作業系統。您可以取代指令碼中的 CUDA 版本，為 Amazon EC2 執行個體中的任何可用 CUDA 版本執行命令。 + Amazon Linux 2、Amazon Linux 2023： ``` $ /opt/amazon/openmpi/bin/mpirun -n 16 -N 8 \ -x NCCL_DEBUG=INFO --mca pml ^cm \ -x LD_LIBRARY_PATH=/usr/local/cuda-xx.x/efa/lib:/usr/local/cuda-xx.x/lib:/usr/local/cuda-xx.x/lib64:/usr/local/cuda-xx.x:/opt/amazon/efa/lib64:/opt/amazon/openmpi/lib64:$LD_LIBRARY_PATH \ --hostfile hosts --mca btl tcp,self --mca btl_tcp_if_exclude lo,docker0 --bind-to none \ /usr/local/cuda-xx.x/efa/test-cuda-xx.x/all_reduce_perf -b 8 -e 1G -f 2 -g 1 -c 1 -n 100 | tee ${TRAINING_LOG} ``` + Ubuntu 20.04、Ubuntu 20.04： ``` $ /opt/amazon/openmpi/bin/mpirun -n 16 -N 8 \ -x NCCL_DEBUG=INFO --mca pml ^cm \ -x LD_LIBRARY_PATH=/usr/local/cuda-xx.x/efa/lib:/usr/local/cuda-xx.x/lib:/usr/local/cuda-xx.x/lib64:/usr/local/cuda-xx.x:/opt/amazon/efa/lib:/opt/amazon/openmpi/lib:$LD_LIBRARY_PATH \ --hostfile hosts --mca btl tcp,self --mca btl_tcp_if_exclude lo,docker0 --bind-to none \ /usr/local/cuda-xx.x/efa/test-cuda-xx.x/all_reduce_perf -b 8 -e 1G -f 2 -g 1 -c 1 -n 100 | tee ${TRAINING_LOG} ``` 您的輸出看起來應如以下所示： ``` # nThread 1 nGpus 1 minBytes 8 maxBytes 1073741824 step: 2(factor) warmup iters: 5 iters: 100 agg iters: 1 validation: 1 graph: 0 # # Using devices # Rank 0 Group 0 Pid 33378 on ip-172-31-42-25 device 0 [0x10] NVIDIA A100-SXM4-40GB # Rank 1 Group 0 Pid 33379 on ip-172-31-42-25 device 1 [0x10] NVIDIA A100-SXM4-40GB # Rank 2 Group 0 Pid 33380 on ip-172-31-42-25 device 2 [0x20] NVIDIA A100-SXM4-40GB # Rank 3 Group 0 Pid 33381 on ip-172-31-42-25 device 3 [0x20] NVIDIA A100-SXM4-40GB # Rank 4 Group 0 Pid 33382 on ip-172-31-42-25 device 4 [0x90] NVIDIA A100-SXM4-40GB # Rank 5 Group 0 Pid 33383 on ip-172-31-42-25 device 5 [0x90] NVIDIA A100-SXM4-40GB # Rank 6 Group 0 Pid 33384 on ip-172-31-42-25 device 6 [0xa0] NVIDIA A100-SXM4-40GB # Rank 7 Group 0 Pid 33385 on ip-172-31-42-25 device 7 [0xa0] NVIDIA A100-SXM4-40GB # Rank 8 Group 0 Pid 30378 on ip-172-31-43-8 device 0 [0x10] NVIDIA A100-SXM4-40GB # Rank 9 Group 0 Pid 30379 on ip-172-31-43-8 device 1 [0x10] NVIDIA A100-SXM4-40GB # Rank 10 Group 0 Pid 30380 on ip-172-31-43-8 device 2 [0x20] NVIDIA A100-SXM4-40GB # Rank 11 Group 0 Pid 30381 on ip-172-31-43-8 device 3 [0x20] NVIDIA A100-SXM4-40GB # Rank 12 Group 0 Pid 30382 on ip-172-31-43-8 device 4 [0x90] NVIDIA A100-SXM4-40GB # Rank 13 Group 0 Pid 30383 on ip-172-31-43-8 device 5 [0x90] NVIDIA A100-SXM4-40GB # Rank 14 Group 0 Pid 30384 on ip-172-31-43-8 device 6 [0xa0] NVIDIA A100-SXM4-40GB # Rank 15 Group 0 Pid 30385 on ip-172-31-43-8 device 7 [0xa0] NVIDIA A100-SXM4-40GB ip-172-31-42-25:33385:33385 [7] NCCL INFO cudaDriverVersion 12060 ip-172-31-43-8:30383:30383 [5] NCCL INFO Bootstrap : Using ens32:172.31.43.8 ip-172-31-43-8:30383:30383 [5] NCCL INFO NCCL version 2.23.4+cuda12.5 ... ip-172-31-42-25:33384:33451 [6] NCCL INFO NET/OFI Initializing aws-ofi-nccl 1.13.2-aws ip-172-31-42-25:33384:33451 [6] NCCL INFO NET/OFI Using Libfabric version 1.22 ip-172-31-42-25:33384:33451 [6] NCCL INFO NET/OFI Using CUDA driver version 12060 with runtime 12050 ip-172-31-42-25:33384:33451 [6] NCCL INFO NET/OFI Configuring AWS-specific options ip-172-31-42-25:33384:33451 [6] NCCL INFO NET/OFI Setting provider_filter to efa ip-172-31-42-25:33384:33451 [6] NCCL INFO NET/OFI Setting FI_EFA_FORK_SAFE environment variable to 1 ip-172-31-42-25:33384:33451 [6] NCCL INFO NET/OFI Setting NCCL_NVLSTREE_MAX_CHUNKSIZE to 512KiB ip-172-31-42-25:33384:33451 [6] NCCL INFO NET/OFI Setting NCCL_NVLS_CHUNKSIZE to 512KiB ip-172-31-42-25:33384:33451 [6] NCCL INFO NET/OFI Running on p4d.24xlarge platform, Setting NCCL_TOPO_FILE environment variable to /opt/amazon/ofi-nccl/share/aws-ofi-nccl/xml/p4d-24xl-topo.xml ... -----------------------------some output truncated----------------------------------- # out-of-place in-place # size count type redop root time algbw busbw #wrong time algbw busbw #wrong # (B) (elements) (us) (GB/s) (GB/s) (us) (GB/s) (GB/s) 8 2 float sum -1 180.3 0.00 0.00 0 179.3 0.00 0.00 0 16 4 float sum -1 178.1 0.00 0.00 0 177.6 0.00 0.00 0 32 8 float sum -1 178.5 0.00 0.00 0 177.9 0.00 0.00 0 64 16 float sum -1 178.8 0.00 0.00 0 178.7 0.00 0.00 0 128 32 float sum -1 178.2 0.00 0.00 0 177.8 0.00 0.00 0 256 64 float sum -1 178.6 0.00 0.00 0 178.8 0.00 0.00 0 512 128 float sum -1 177.2 0.00 0.01 0 177.1 0.00 0.01 0 1024 256 float sum -1 179.2 0.01 0.01 0 179.3 0.01 0.01 0 2048 512 float sum -1 181.3 0.01 0.02 0 181.2 0.01 0.02 0 4096 1024 float sum -1 184.2 0.02 0.04 0 183.9 0.02 0.04 0 8192 2048 float sum -1 191.2 0.04 0.08 0 190.6 0.04 0.08 0 16384 4096 float sum -1 202.5 0.08 0.15 0 202.3 0.08 0.15 0 32768 8192 float sum -1 233.0 0.14 0.26 0 232.1 0.14 0.26 0 65536 16384 float sum -1 238.6 0.27 0.51 0 235.1 0.28 0.52 0 131072 32768 float sum -1 237.2 0.55 1.04 0 236.8 0.55 1.04 0 262144 65536 float sum -1 248.3 1.06 1.98 0 247.0 1.06 1.99 0 524288 131072 float sum -1 309.2 1.70 3.18 0 307.7 1.70 3.20 0 1048576 262144 float sum -1 408.7 2.57 4.81 0 404.3 2.59 4.86 0 2097152 524288 float sum -1 613.5 3.42 6.41 0 607.9 3.45 6.47 0 4194304 1048576 float sum -1 924.5 4.54 8.51 0 914.8 4.58 8.60 0 8388608 2097152 float sum -1 1059.5 7.92 14.85 0 1054.3 7.96 14.92 0 16777216 4194304 float sum -1 1269.9 13.21 24.77 0 1272.0 13.19 24.73 0 33554432 8388608 float sum -1 1642.7 20.43 38.30 0 1636.7 20.50 38.44 0 67108864 16777216 float sum -1 2446.7 27.43 51.43 0 2445.8 27.44 51.45 0 134217728 33554432 float sum -1 4143.6 32.39 60.73 0 4142.4 32.40 60.75 0 268435456 67108864 float sum -1 7351.9 36.51 68.46 0 7346.7 36.54 68.51 0 536870912 134217728 float sum -1 13717 39.14 73.39 0 13703 39.18 73.46 0 1073741824 268435456 float sum -1 26416 40.65 76.21 0 26420 40.64 76.20 0 ... # Out of bounds values : 0 OK # Avg bus bandwidth : 15.5514 ``` #### 驗證測試 若要驗證 EFA 測試是否傳回有效的結果，請使用下列測試來確認： + 使用 EC2 執行個體中繼資料取得執行個體類型： ``` TOKEN=$(curl -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 21600") INSTANCE_TYPE=$(curl -H "X-aws-ec2-metadata-token: $TOKEN" -v http://169.254.169.254/latest/meta-data/instance-type) ``` + 執行 [效能測試](#tutorial-efa-using-multinode) + 設定下列參數 ``` CUDA_VERSION CUDA_RUNTIME_VERSION NCCL_VERSION ``` + 驗證結果，如下所示： ``` RETURN_VAL=`echo $?` if [ ${RETURN_VAL} -eq 0 ]; then # [0] NCCL INFO NET/OFI Initializing aws-ofi-nccl 1.13.2-aws # [0] NCCL INFO NET/OFI Using CUDA driver version 12060 with runtime 12010 # cudaDriverVersion 12060 --> This is max supported cuda version by nvidia driver # NCCL version 2.23.4+cuda12.5 --> This is NCCL version compiled with cuda version # Validation of logs grep "NET/OFI Configuring AWS-specific options" ${TRAINING_LOG} || { echo "AWS-specific options text not found"; exit 1; } grep "busbw" ${TRAINING_LOG} || { echo "busbw text not found"; exit 1; } grep "Avg bus bandwidth " ${TRAINING_LOG} || { echo "Avg bus bandwidth text not found"; exit 1; } grep "NCCL version $NCCL_VERSION" ${TRAINING_LOG} || { echo "Text not found: NCCL version $NCCL_VERSION"; exit 1; } if [[ ${INSTANCE_TYPE} == "p4d.24xlarge" ]]; then grep "NET/Libfabric/0/GDRDMA" ${TRAINING_LOG} || { echo "Text not found: NET/Libfabric/0/GDRDMA"; exit 1; } grep "NET/OFI Selected Provider is efa (found 4 nics)" ${TRAINING_LOG} || { echo "Selected Provider is efa text not found"; exit 1; } elif [[ ${INSTANCE_TYPE} == "p4de.24xlarge" ]]; then grep "NET/Libfabric/0/GDRDMA" ${TRAINING_LOG} || { echo "Avg bus bandwidth text not found"; exit 1; } grep "NET/OFI Selected Provider is efa (found 4 nics)" ${TRAINING_LOG} || { echo "Avg bus bandwidth text not found"; exit 1; } elif [[ ${INSTANCE_TYPE} == "p5.48xlarge" ]]; then grep "NET/Libfabric/0/GDRDMA" ${TRAINING_LOG} || { echo "Avg bus bandwidth text not found"; exit 1; } grep "NET/OFI Selected Provider is efa (found 32 nics)" ${TRAINING_LOG} || { echo "Avg bus bandwidth text not found"; exit 1; } elif [[ ${INSTANCE_TYPE} == "p5e.48xlarge" ]]; then grep "NET/Libfabric/0/GDRDMA" ${TRAINING_LOG} || { echo "Avg bus bandwidth text not found"; exit 1; } grep "NET/OFI Selected Provider is efa (found 32 nics)" ${TRAINING_LOG} || { echo "Avg bus bandwidth text not found"; exit 1; } elif [[ ${INSTANCE_TYPE} == "p5en.48xlarge" ]]; then grep "NET/Libfabric/0/GDRDMA" ${TRAINING_LOG} || { echo "Avg bus bandwidth text not found"; exit 1; } grep "NET/OFI Selected Provider is efa (found 16 nics)" ${TRAINING_LOG} || { echo "Avg bus bandwidth text not found"; exit 1; } elif [[ ${INSTANCE_TYPE} == "p3dn.24xlarge" ]]; then grep "NET/OFI Selected Provider is efa (found 4 nics)" ${TRAINING_LOG} || { echo "Selected Provider is efa text not found"; exit 1; } fi echo "***************************** check_efa_nccl_all_reduce passed for cuda version ${CUDA_VERSION} *****************************" else echo "***************************** check_efa_nccl_all_reduce failed for cuda version ${CUDA_VERSION} *****************************" fi ``` + 若要存取基準資料，我們可以剖析多節點 all\$1reduce 測試的資料表輸出最後一列： ``` benchmark=$(sudo cat ${TRAINING_LOG} | grep '1073741824' | tail -n1 | awk -F " " '{{print $12}}' | sed 's/ //' | sed 's/ 5e-07//') if [[ -z "${benchmark}" ]]; then echo "benchmark variable is empty" exit 1 fi echo "Benchmark throughput: ${benchmark}" ```