What is NCCL
NCCL (NVIDIA Collective Communications Library) 是 NVIDIA 推出的一个用于 GPU 之间高性能通信的库。随着深度学习模型规模的增长(如 GPT-3 的 1750 亿参数),单个 GPU 已无法满足训练需求。这就需要将模型或数据分割到多个 GPU 上进行并行训练,而 GPU 之间必然需要进行数据交换。NCCL 就是为了解决这个场景而生的。它主要解决以下问题:
在多 GPU 训练场景下实现高效的数据交换
自动识别并优化 GPU 间的通信拓扑
提供标准化的集合通信接口
支持机内(单机多卡)和机间(多机多卡)通信
NCCL 重要概念
Collective Operations:NCCL 支持各种集体通信操作,如广播(Broadcast)、规约(Reduction)、聚合(Aggregation)、AllReduce、AllGather 等。这些操作可以在多个 GPU 或节点之间进行数据同步和合并。
Processes and Groups:NCCL 中的进程(Processes)表示参与通信的计算节点。进程可以组织成组(Groups),以便在组内进行集体通信。
Communicators:Communicators 是 NCCL 中用于定义进程之间通信关系的对象。它指定了参与通信的进程组和通信模式(如点对点、广播等)。
Devices and Streams:NCCL 与 GPU 设备密切相关,它可以在多个 GPU 设备之间进行通信。同时,NCCL 还支持与 CUDA 流(Streams)的集成,以实现异步通信和并行计算。
Synchronization:在分布式计算中,同步是至关重要的。NCCL 提供了各种同步原语,如 barrier 同步,以确保进程在执行集体通信操作时达到一致的状态。
Performance Optimization:NCCL 注重性能优化,它提供了一些技术来提高通信效率,如集体通信的合并、数据传输的批量处理、通信与计算的重叠等。
Fault Tolerance:NCCL 还考虑了容错性,支持在部分节点故障或网络不稳定的情况下进行可靠的通信。
NCCL 支持的操作类型
NCCL 支持以下几种主要的集合通信操作:
AllReduce:所有 GPU 的数据先进行规约(如求和),然后广播到所有 GPU
Broadcast:从一个源 GPU 向其他所有 GPU 广播数据
Reduce:将所有 GPU 的数据规约到一个目标 GPU
AllGather: 收集所有 GPU 的数据并分发给所有 GPU
ReduceScatter: 规约后将结果分散到所有 GPU
Send/Recv: 点对点通信
AllToAll: 将数据分发到所有 GPU
NCCL 的设备要求
NVIDIA GPU (支持 CUDA)
推荐使用支持 NVLink 的 GPU 以获得最佳性能
对于多机通信,建议使用 InfiniBand 或 RoCE 网络
需要安装对应版本的 CUDA 和 GPU 驱动
NCCL in PyTorch
PyTorch 内置 NCCL 后端支持,使用非常简单:
import torch.distributed as dist# 初始化进程组dist.init_process_group(backend='nccl')# 创建分布式模型model = torch.nn.parallel.DistributedDataParallel(model)
NCCL 的工作方式
Ring Algorithm
NCCL 在机内通信时主要使用 Ring Algorithm。其核心思想是:
将所有 GPU 组织成一个环
每个 GPU 只与其相邻的 GPU 通信
通过多轮次传递实现全局数据交换
数据被分成多个 chunk 进行并行传输
优点:- 充分利用 GPU 间的带宽 - 通信负载均衡 - 易于扩展
Tree Algorithm
在torch-distributed 的后记中已经介绍过了。
通信协议
NCCL 实现了三种通信协议:
Simple: 基础协议
LL(Low Latency): 低延迟协议,适用于小数据量
LL128: 在 NVLink 环境下的优化协议,可达到 93.75% 的有效带宽
NCCL 与其他通信库的对比
与 MPI 的区别:
NCCL 专注于 GPU 通信优化
MPI 更通用但性能可能较低
可以结合使用:MPI 管理进程,NCCL 负责 GPU 通信
与 Gloo 的区别:
Gloo 支持 CPU 和 GPU
NCCL 在 GPU 通信性能上更优
常见 NIVIDA 指令
这里参考了 WeLearnNLP 的指南(https://www.yourmetaverse.cn/deep_learning/199/)。
nvidia-smi topo -m
最典型的当然有 nvidia-smi
和 nvidia-smi topo -m
。前者都非常熟悉了,这里我对比下两台集群的 nvidia-smi topo -m
的输出:
GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7 CPU Affinity NUMA Affinity GPU NUMA ID
GPU0 X SYS SYS SYS SYS SYS SYS SYS 0-15,32-47 0 N/A
GPU1 SYS X SYS SYS SYS SYS SYS SYS 0-15,32-47 0 N/A
GPU2 SYS SYS X SYS SYS SYS SYS SYS 0-15,32-47 0 N/A
GPU3 SYS SYS SYS X SYS SYS SYS SYS 0-15,32-47 0 N/A
GPU4 SYS SYS SYS SYS X SYS SYS SYS 16-31,48-63 1 N/A
GPU5 SYS SYS SYS SYS SYS X SYS SYS 16-31,48-63 1 N/A
GPU6 SYS SYS SYS SYS SYS SYS X SYS 16-31,48-63 1 N/A
GPU7 SYS SYS SYS SYS SYS SYS SYS X 16-31,48-63 1 N/A
Legend: X = Self SYS = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node PHB = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
PXB = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
PIX = Connection traversing at most a single PCIe bridge
NV# = Connection traversing a bonded set of # NVLinksGPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7 CPU Affinity NUMA Affinity GPU NUMA ID
GPU0 X NV18 NV18 NV18 NV18 NV18 NV18 NV18 0-47,96-143 0 N/A
GPU1 NV18 X NV18 NV18 NV18 NV18 NV18 NV18 0-47,96-143 0 N/A
GPU2 NV18 NV18 X NV18 NV18 NV18 NV18 NV18 0-47,96-143 0 N/A
GPU3 NV18 NV18 NV18 X NV18 NV18 NV18 NV18 0-47,96-143 0 N/A
GPU4 NV18 NV18 NV18 NV18 X NV18 NV18 NV18 48-95,144-191 1 N/A
GPU5 NV18 NV18 NV18 NV18 NV18 X NV18 NV18 48-95,144-191 1 N/A
GPU6 NV18 NV18 NV18 NV18 NV18 NV18 X NV18 48-95,144-191 1 N/A
GPU7 NV18 NV18 NV18 NV18 NV18 NV18 NV18 X 48-95,144-191 1 N/A
Legend: X = Self SYS = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node PHB = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
PXB = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
PIX = Connection traversing at most a single PCIe bridge
NV# = Connection traversing a bonded set of # NVLinks
可以读出很多有趣的信息:
通过对比这两个集群的拓扑信息,我可以得出以下几个重要结论:
互联方式
第一个集群:所有 GPU 之间通过 PCIe 和 NUMA 节点间的 SMP 互联(标记为 SYS)
第二个集群:所有 GPU 之间通过 18 条 NVLink 连接(标记为 NV18)
性能影响:第二个集群的 GPU 间通信性能显著优于第一个集群,因为 NVLink 的带宽和延迟都优于 PCIe+SMP 方案
NUMA 架构
两个集群都采用双 NUMA 节点设计:
GPU 0-3 属于 NUMA 节点 0
GPU 4-7 属于 NUMA 节点 1
GPU 通信:应尽量将相关任务分配到同一 NUMA 节点内的 GPU,以避免跨 NUMA 节点的频繁数据传输
CPU 核心分配:
第一个集群:每个 NUMA 节点分配 32 个核心(如 0-15,32-47)
第二个集群:每个 NUMA 节点分配 96 个核心(如 0-47,96-143)
系统规模
GPU 数量:两个集群都是 8 GPU 配置
CPU 核心总数:
第一个集群:64 核心
第二个集群:192 核心
拓扑完整性
每个 GPU 都与其他所有 GPU 直接相连
NVLINK 查询
nvidia-smi nvlink --status -i 0nvidia-smi nvlink --capabilities-i 0
nvlink 查询结果
GPU 0: NVIDIA H100 80GB HBM3 (UUID: GPU-5a10e6e5-95f7-2785-ed63-6f6147f304f7)
Link 0: 26.562 GB/s
Link 1: 26.562 GB/s
Link 2: 26.562 GB/s
Link 3: 26.562 GB/s
Link 4: 26.562 GB/s
Link 5: 26.562 GB/s
Link 6: 26.562 GB/s
Link 7: 26.562 GB/s
Link 8: 26.562 GB/s
Link 9: 26.562 GB/s
Link 10: 26.562 GB/s
Link 11: 26.562 GB/s
Link 12: 26.562 GB/s
Link 13: 26.562 GB/s
Link 14: 26.562 GB/s
Link 15: 26.562 GB/s
Link 16: 26.562 GB/s
Link 17: 26.562 GB/s
GPU 0: NVIDIA H100 80GB HBM3 (UUID: GPU-5a10e6e5-95f7-2785-ed63-6f6147f304f7)
Link 0, P2P is supported: true
Link 0, Access to system memory supported: true
Link 0, P2P atomics supported: true
Link 0, System memory atomics supported: true
Link 0, SLI is supported: true
Link 0, Link is supported: true
Link 1, P2P is supported: true
Link 1, Access to system memory supported: true
Link 1, P2P atomics supported: true
Link 1, System memory atomics supported: true
Link 1, SLI is supported: true
Link 1, Link is supported: true
Link 2, P2P is supported: true
Link 2, Access to system memory supported: true
Link 2, P2P atomics supported: true
Link 2, System memory atomics supported: true
Link 2, SLI is supported: true
Link 2, Link is supported: true
Link 3, P2P is supported: true
Link 3, Access to system memory supported: true
Link 3, P2P atomics supported: true
Link 3, System memory atomics supported: true
Link 3, SLI is supported: true
Link 3, Link is supported: true
Link 4, P2P is supported: true
Link 4, Access to system memory supported: true
Link 4, P2P atomics supported: true
Link 4, System memory atomics supported: true
Link 4, SLI is supported: true
Link 4, Link is supported: true
Link 5, P2P is supported: true
Link 5, Access to system memory supported: true
Link 5, P2P atomics supported: true
Link 5, System memory atomics supported: true
Link 5, SLI is supported: true
Link 5, Link is supported: true
Link 6, P2P is supported: true
Link 6, Access to system memory supported: true
Link 6, P2P atomics supported: true
Link 6, System memory atomics supported: true
Link 6, SLI is supported: true
Link 6, Link is supported: true
Link 7, P2P is supported: true
Link 7, Access to system memory supported: true
Link 7, P2P atomics supported: true
Link 7, System memory atomics supported: true
Link 7, SLI is supported: true
Link 7, Link is supported: true
Link 8, P2P is supported: true
Link 8, Access to system memory supported: true
Link 8, P2P atomics supported: true
Link 8, System memory atomics supported: true
Link 8, SLI is supported: true
Link 8, Link is supported: true
Link 9, P2P is supported: true
Link 9, Access to system memory supported: true
Link 9, P2P atomics supported: true
Link 9, System memory atomics supported: true
Link 9, SLI is supported: true
Link 9, Link is supported: true
Link 10, P2P is supported: true
Link 10, Access to system memory supported: true
Link 10, P2P atomics supported: true
Link 10, System memory atomics supported: true
Link 10, SLI is supported: true
Link 10, Link is supported: true
Link 11, P2P is supported: true
Link 11, Access to system memory supported: true
Link 11, P2P atomics supported: true
Link 11, System memory atomics supported: true
Link 11, SLI is supported: true
Link 11, Link is supported: true
Link 12, P2P is supported: true
Link 12, Access to system memory supported: true
Link 12, P2P atomics supported: true
Link 12, System memory atomics supported: true
Link 12, SLI is supported: true
Link 12, Link is supported: true
Link 13, P2P is supported: true
Link 13, Access to system memory supported: true
Link 13, P2P atomics supported: true
Link 13, System memory atomics supported: true
Link 13, SLI is supported: true
Link 13, Link is supported: true
Link 14, P2P is supported: true
Link 14, Access to system memory supported: true
Link 14, P2P atomics supported: true
Link 14, System memory atomics supported: true
Link 14, SLI is supported: true
Link 14, Link is supported: true
Link 15, P2P is supported: true
Link 15, Access to system memory supported: true
Link 15, P2P atomics supported: true
Link 15, System memory atomics supported: true
Link 15, SLI is supported: true
Link 15, Link is supported: true
Link 16, P2P is supported: true
Link 16, Access to system memory supported: true
Link 16, P2P atomics supported: true
Link 16, System memory atomics supported: true
Link 16, SLI is supported: true
Link 16, Link is supported: true
Link 17, P2P is supported: true
Link 17, Access to system memory supported: true
Link 17, P2P atomics supported: true
Link 17, System memory atomics supported: true
Link 17, SLI is supported: true
Link 17, Link is supported: true
可以分析看到一些对开发实用的特性:
P2P(点对点)通信
系统内存访问
P2P原子操作
系统内存原子操作
SLI(多GPU并行)
完整的链路支持
GPU 监控
可以监控 GPU 的方式很多,这里推荐 nvitop(https://github.com/Syllo/nvtop),非常方便,pip 安装即可,看着最赏心悦目。
- The End -
长按二维码关注我们
本公众号专注:
1. 技术分享;
2. 学术交流;
3. 资料共享。
欢迎关注我们,一起成长!