集合通信行为分析 基于NCCL

姚伟峰

• 集合通信行为分析 - 基于NCCL
  • Phases
  • NCCL
    • Topology Phase
      • Build Physical Topology (i.e. System Topology)
        • Transport Types
      • Build Logical Topology (i.e. Graph Topology)
        • Tree Logical Topology
          • log
          • 解析
          • 结果
        • Ring Logical Topology
          • log
          • 解析
          • 结果
    • Collective Phase
  • Reference

Phases

from
Phase 1 - Bootstrap Phase: Initiate all nodes and then all ranks in a collective. It makes sure all ranks know about all other ranks, so any rank is able to communicate with any other rank.
Phase 2 - Topology Phase: Each node detects and maps out what hardware is located on the machine. Hardware includes CPUs, GPUs, NICs and interconnect types. Each node then creates an intra-machine graph, connects hardware with PCIe or NVLink interconnect, and evaluates the graph. When the intra-machine topology is decided, the system will decide what pattern to use for the whole system. The two main patterns are a tree or a ring. While the topology is evaluated, NCCL is also tuning it by performing tests. This allows each rank to pre-compute thresholds for message sizes.
Phase 3 - Collective Phase: A user can dispatch many collective operations using the same topology.

NCCL

Topology Phase

Build Physical Topology (i.e. System Topology)

建立rank间的邻接表。

Transport Types

查看代码

#define TRANSPORT_P2P 0
#define TRANSPORT_SHM 1
#define TRANSPORT_NET 2

Build Logical Topology (i.e. Graph Topology)

以2机16卡, NCCL 2.8.4为例

NCCL会构建tree，ring graph。

Tree Logical Topology

log

查看代码

10.0.2.11: 2be7fa6883db:57976:58906 [5] NCCL INFO Trees [0] 14/-1/-1->13->12 [1] 14/-1/-1->13->12
10.0.2.11: 2be7fa6883db:57977:58920 [6] NCCL INFO Trees [0] 15/-1/-1->14->13 [1] 15/-1/-1->14->13
10.0.2.11: 2be7fa6883db:57978:58913 [7] NCCL INFO Trees [0] -1/-1/-1->15->14 [1] -1/-1/-1->15->14
10.0.2.11: 2be7fa6883db:57975:58907 [4] NCCL INFO Trees [0] 13/-1/-1->12->11 [1] 13/-1/-1->12->11
10.0.2.11: 2be7fa6883db:57974:58908 [3] NCCL INFO Trees [0] 12/-1/-1->11->10 [1] 12/-1/-1->11->10
10.0.2.11: 2be7fa6883db:57971:58905 [0] NCCL INFO Trees [0] 9/-1/-1->8->0 [1] 9/0/-1->8->-1
10.0.2.11: 2be7fa6883db:57973:58909 [2] NCCL INFO Trees [0] 11/-1/-1->10->9 [1] 11/-1/-1->10->9
10.0.2.11: 2be7fa6883db:57972:58904 [1] NCCL INFO Trees [0] 10/-1/-1->9->8 [1] 10/-1/-1->9->8
10.0.2.12: 94f182076445:82266:83142 [5] NCCL INFO Trees [0] 6/-1/-1->5->4 [1] 6/-1/-1->5->4
10.0.2.12: 94f182076445:82263:83145 [2] NCCL INFO Trees [0] 3/-1/-1->2->1 [1] 3/-1/-1->2->1
10.0.2.12: 94f182076445:82262:83144 [1] NCCL INFO Trees [0] 2/-1/-1->1->0 [1] 2/-1/-1->1->0
10.0.2.12: 94f182076445:82267:83151 [6] NCCL INFO Trees [0] 7/-1/-1->6->5 [1] 7/-1/-1->6->5
10.0.2.12: 94f182076445:82265:83143 [4] NCCL INFO Trees [0] 5/-1/-1->4->3 [1] 5/-1/-1->4->3
10.0.2.12: 94f182076445:82261:83141 [0] NCCL INFO Trees [0] 1/8/-1->0->-1 [1] 1/-1/-1->0->8
10.0.2.12: 94f182076445:82268:83149 [7] NCCL INFO Trees [0] -1/-1/-1->7->6 [1] -1/-1/-1->7->6
10.0.2.12: 94f182076445:82264:83150 [3] NCCL INFO Trees [0] 4/-1/-1->3->2 [1] 4/-1/-1->3->2
10.0.2.12: 94f182076445:82263:83145 [2] NCCL INFO Channel 00 : 2[1d000] -> 1[1b000] via P2P/IPC
10.0.2.12: 94f182076445:82265:83143 [4] NCCL INFO Channel 00 : 4[3d000] -> 3[1e000] via P2P/IPC
10.0.2.12: 94f182076445:82264:83150 [3] NCCL INFO Channel 00 : 3[1e000] -> 2[1d000] via P2P/IPC
10.0.2.12: 94f182076445:82262:83144 [1] NCCL INFO Channel 00 : 1[1b000] -> 0[1a000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57976:58906 [5] NCCL INFO Channel 00 : 13[3e000] -> 12[3d000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57973:58909 [2] NCCL INFO Channel 00 : 10[1d000] -> 9[1b000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57975:58907 [4] NCCL INFO Channel 00 : 12[3d000] -> 11[1e000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57972:58904 [1] NCCL INFO Channel 00 : 9[1b000] -> 8[1a000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57974:58908 [3] NCCL INFO Channel 00 : 11[1e000] -> 10[1d000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57977:58920 [6] NCCL INFO Channel 00 : 14[40000] -> 13[3e000] via P2P/IPC
10.0.2.12: 94f182076445:82261:83141 [0] NCCL INFO Channel 00 : 8[1a000] -> 0[1a000] [receive] via NET/Socket/0
10.0.2.11: 2be7fa6883db:57978:58913 [7] NCCL INFO Channel 00 : 15[41000] -> 14[40000] via P2P/IPC
10.0.2.12: 94f182076445:82267:83151 [6] NCCL INFO Channel 00 : 6[40000] -> 5[3e000] via P2P/IPC
10.0.2.12: 94f182076445:82266:83142 [5] NCCL INFO Channel 00 : 5[3e000] -> 4[3d000] via P2P/IPC
10.0.2.12: 94f182076445:82261:83141 [0] NCCL INFO Channel 00 : 0[1a000] -> 8[1a000] [send] via NET/Socket/0
10.0.2.12: 94f182076445:82268:83149 [7] NCCL INFO Channel 00 : 7[41000] -> 6[40000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57971:58905 [0] NCCL INFO Channel 00 : 0[1a000] -> 8[1a000] [receive] via NET/Socket/0
10.0.2.11: 2be7fa6883db:57971:58905 [0] NCCL INFO Channel 00 : 8[1a000] -> 0[1a000] [send] via NET/Socket/0
10.0.2.12: 94f182076445:82263:83145 [2] NCCL INFO Channel 01 : 2[1d000] -> 1[1b000] via P2P/IPC
10.0.2.12: 94f182076445:82265:83143 [4] NCCL INFO Channel 01 : 4[3d000] -> 3[1e000] via P2P/IPC
10.0.2.12: 94f182076445:82264:83150 [3] NCCL INFO Channel 01 : 3[1e000] -> 2[1d000] via P2P/IPC
10.0.2.12: 94f182076445:82262:83144 [1] NCCL INFO Channel 01 : 1[1b000] -> 0[1a000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57976:58906 [5] NCCL INFO Channel 01 : 13[3e000] -> 12[3d000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57973:58909 [2] NCCL INFO Channel 01 : 10[1d000] -> 9[1b000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57975:58907 [4] NCCL INFO Channel 01 : 12[3d000] -> 11[1e000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57972:58904 [1] NCCL INFO Channel 01 : 9[1b000] -> 8[1a000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57974:58908 [3] NCCL INFO Channel 01 : 11[1e000] -> 10[1d000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57977:58920 [6] NCCL INFO Channel 01 : 14[40000] -> 13[3e000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57978:58913 [7] NCCL INFO Channel 01 : 15[41000] -> 14[40000] via P2P/IPC
10.0.2.12: 94f182076445:82267:83151 [6] NCCL INFO Channel 01 : 6[40000] -> 5[3e000] via P2P/IPC
10.0.2.12: 94f182076445:82266:83142 [5] NCCL INFO Channel 01 : 5[3e000] -> 4[3d000] via P2P/IPC
10.0.2.12: 94f182076445:82261:83141 [0] NCCL INFO Channel 01 : 8[1a000] -> 0[1a000] [receive] via NET/Socket/0
10.0.2.12: 94f182076445:82261:83141 [0] NCCL INFO Channel 01 : 0[1a000] -> 8[1a000] [send] via NET/Socket/0
10.0.2.12: 94f182076445:82268:83149 [7] NCCL INFO Channel 01 : 7[41000] -> 6[40000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57971:58905 [0] NCCL INFO Channel 01 : 0[1a000] -> 8[1a000] [receive] via NET/Socket/0
10.0.2.11: 2be7fa6883db:57971:58905 [0] NCCL INFO Channel 01 : 8[1a000] -> 0[1a000] [send] via NET/Socket/0

解析

• 拓扑log格式

  IP: hostname:pid:tid [cudaDev] NCCL INFO Trees [channel ID] down0 rank/down1 rank/down2 rank->current rank->up rank
  如下面log中：
  10.0.2.11: 2be7fa6883db:57976:58906 [5] NCCL INFO Trees [0] 14/-1/-1->13->12 [1] 14/-1/-1->13->12
  可以解读为 10.0.2.11上的设备5，其rank为13，有两棵树，分别为channel 0和channel 1: channel 0的子节点只有14, 父节点为12; channel 1一样。

• channel log格式

  IP: hostname:pid:tid [cudaDev] NCCL INFO Channel [channel ID] current rank[bus ID]->successor rank[bus ID] via transport type
  如下面log中：
  10.0.2.11: 2be7fa6883db:57976:58906 [5] NCCL INFO Channel 00 : 13[3e000] -> 14[40000] via P2P/IPC
  可以解读为 10.0.2.11上的设备5(rank 为13, bus ID为3e000)，其channel 0连接至rank 14，传输方式为P2P/IPC。

结果

依此解析，可得两棵一样的tree，逻辑拓扑如下：

其中socket双工通道建立如下(双工为1个channel)：

Ring Logical Topology

log

查看代码

10.0.2.12: 94f182076445:82261:83141 [0] NCCL INFO Channel 00/02 : 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
10.0.2.12: 94f182076445:82261:83141 [0] NCCL INFO Channel 01/02 : 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
10.0.2.12: 94f182076445:82263:83145 [2] NCCL INFO Channel 00 : 2[1d000] -> 3[1e000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57978:58913 [7] NCCL INFO Channel 00 : 15[41000] -> 0[1a000] [send] via NET/Socket/0
10.0.2.12: 94f182076445:82261:83141 [0] NCCL INFO Channel 00 : 15[41000] -> 0[1a000] [receive] via NET/Socket/0
10.0.2.12: 94f182076445:82266:83142 [5] NCCL INFO Channel 00 : 5[3e000] -> 6[40000] via P2P/IPC
10.0.2.12: 94f182076445:82265:83143 [4] NCCL INFO Channel 00 : 4[3d000] -> 5[3e000] via P2P/IPC
10.0.2.12: 94f182076445:82262:83144 [1] NCCL INFO Channel 00 : 1[1b000] -> 2[1d000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57976:58906 [5] NCCL INFO Channel 00 : 13[3e000] -> 14[40000] via P2P/IPC
10.0.2.12: 94f182076445:82264:83150 [3] NCCL INFO Channel 00 : 3[1e000] -> 4[3d000] via P2P/IPC
10.0.2.12: 94f182076445:82267:83151 [6] NCCL INFO Channel 00 : 6[40000] -> 7[41000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57977:58920 [6] NCCL INFO Channel 00 : 14[40000] -> 15[41000] via P2P/IPC
10.0.2.12: 94f182076445:82268:83149 [7] NCCL INFO Channel 00 : 7[41000] -> 8[1a000] [send] via NET/Socket/0
10.0.2.11: 2be7fa6883db:57971:58905 [0] NCCL INFO Channel 00 : 7[41000] -> 8[1a000] [receive] via NET/Socket/0
10.0.2.12: 94f182076445:82261:83141 [0] NCCL INFO Channel 00 : 0[1a000] -> 1[1b000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57972:58904 [1] NCCL INFO Channel 00 : 9[1b000] -> 10[1d000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57974:58908 [3] NCCL INFO Channel 00 : 11[1e000] -> 12[3d000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57975:58907 [4] NCCL INFO Channel 00 : 12[3d000] -> 13[3e000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57973:58909 [2] NCCL INFO Channel 00 : 10[1d000] -> 11[1e000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57971:58905 [0] NCCL INFO Channel 00 : 8[1a000] -> 9[1b000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57974:58908 [3] NCCL INFO Channel 01 : 11[1e000] -> 12[3d000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57976:58906 [5] NCCL INFO Channel 01 : 13[3e000] -> 14[40000] via P2P/IPC
10.0.2.12: 94f182076445:82263:83145 [2] NCCL INFO Channel 01 : 2[1d000] -> 3[1e000] via P2P/IPC
10.0.2.12: 94f182076445:82266:83142 [5] NCCL INFO Channel 01 : 5[3e000] -> 6[40000] via P2P/IPC
10.0.2.12: 94f182076445:82265:83143 [4] NCCL INFO Channel 01 : 4[3d000] -> 5[3e000] via P2P/IPC
10.0.2.12: 94f182076445:82262:83144 [1] NCCL INFO Channel 01 : 1[1b000] -> 2[1d000] via P2P/IPC
10.0.2.12: 94f182076445:82264:83150 [3] NCCL INFO Channel 01 : 3[1e000] -> 4[3d000] via P2P/IPC
10.0.2.12: 94f182076445:82267:83151 [6] NCCL INFO Channel 01 : 6[40000] -> 7[41000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57978:58913 [7] NCCL INFO Channel 01 : 15[41000] -> 0[1a000] [send] via NET/Socket/0
10.0.2.12: 94f182076445:82261:83141 [0] NCCL INFO Channel 01 : 15[41000] -> 0[1a000] [receive] via NET/Socket/0
10.0.2.12: 94f182076445:82261:83141 [0] NCCL INFO Channel 01 : 0[1a000] -> 1[1b000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57977:58920 [6] NCCL INFO Channel 01 : 14[40000] -> 15[41000] via P2P/IPC
10.0.2.12: 94f182076445:82268:83149 [7] NCCL INFO Channel 01 : 7[41000] -> 8[1a000] [send] via NET/Socket/0
10.0.2.11: 2be7fa6883db:57972:58904 [1] NCCL INFO Channel 01 : 9[1b000] -> 10[1d000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57975:58907 [4] NCCL INFO Channel 01 : 12[3d000] -> 13[3e000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57973:58909 [2] NCCL INFO Channel 01 : 10[1d000] -> 11[1e000] via P2P/IPC
10.0.2.11: 2be7fa6883db:57971:58905 [0] NCCL INFO Channel 01 : 7[41000] -> 8[1a000] [receive] via NET/Socket/0
10.0.2.11: 2be7fa6883db:57971:58905 [0] NCCL INFO Channel 01 : 8[1a000] -> 9[1b000] via P2P/IPC

解析

• 拓扑log格式

  IP: hostname:pid:tid [cudaDev] NCCL INFO Channel ring_ID/ring_number: rank0 rank1 … last_rank
  如下面log中：
  10.0.2.12: 94f182076445:82261:83141 [0] NCCL INFO Channel 00/02 : 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
  可以解读为：建成了02个ring，其中第0个ring的成员有：0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15，该ring共由16个rank组成。

• channel log格式
  与tree拓扑的格式一致。

结果

依此解析，可得两个一样的ring，逻辑拓扑如下：

Collective Phase

用户调用NCCL支持的集合通信原语进行通信：

• 集合通信原语

  • AllReduce

  • Broadcast

  • Reduce

  • AllGather

  • ReduceScatter

• 点对点通信原语

  • Send

  • Recv

NCCL在getAlgoInfo里面使用ncclTopoGetAlgoTime来遍历计算(algorithm, protocol)，最终选择预测会最快做完指定数据量的指定集合通信原语的algorithm和protocol完成该通信原语。

Reference

1. Breaking the Computation and Communication Abstraction Barrier in Distributed Machine Learning Workloads

2. NNCL github

3. Massively Scale Your Deep Learning Training with NCCL 2.4

4. Distributed Deep Neural Network Training: NCCL on SUMMIT

5. Synthesizing Optimal Collective Algorithms

6. Towards Scalable Distributed Training of Deep Learning on Public Cloud Clusters

7. Fast Multi-GPU communication over PCI Express

8. 腾讯机智团队分享–AllReduce算法的前世今生

9. Fast Multi-GPU collectives with NCCL

10. NCCL: Accelerated Multi-GPU Collective Communications

11. Optimized Broadcast for Deep Learning Workloads on Dense-GPU InfiniBand Clusters: MPI or NCCL?

12. BLINK: Fast and Generic Collectives for Distributed ML

13. MPI Tutorial

14. Optimizing Communication for Clusters of GPUs

15. CNCL Design Docs