aiacc-nccl 2.0.0

AIACC-NCCL is an AI-Accelerator communication framework for NVIDIA-NCCL. It implements optimized all-reduce, all-gather, reduce, broadcast, reduce-scatter, all-to-all,as well as any send/receive based communication pattern.It has been optimized to achieve high bandwidth on aliyun machines using PCIe, NVLink, NVswitch,as well as networking using InfiniBand Verbs, eRDMA or TCP/IP sockets.

AIACC-NCCL

Optimized primitives for inter-GPU communication on Aliyun machines.

Introduction

AIACC-NCCL is an AI-Accelerator communication framework for NVIDIA-NCCL. It implements optimized all-reduce, all-gather, reduce, broadcast, reduce-scatter, all-to-all, as well as any send/receive based communication pattern. It has been optimized to achieve high bandwidth on aliyun machines using PCIe, NVLink, NVswitch, as well as networking using InfiniBand Verbs, eRDMA or TCP/IP sockets.

Install

To install AIACC NCCL on the system, create a package then install it as root as follow two methods:

• method1: rpm/deb (Recommended)

# Centos:
wget http://mirrors.aliyun.com/aiacc/aiacc-nccl/aiacc_nccl-1.0.rpm
rpm -i aiacc-nccl-1.0.rpm
# Ubuntu:
wget http://mirrors.aliyun.com/aiacc/aiacc-nccl/aiacc_nccl-1.0.deb
dpkg -i aiacc-nccl-1.0.deb

• method2: python-offline

wget http://mirrors.aliyun.com/aiacc/aiacc-nccl/aiacc_nccl-2.0.0.tar.gz
pip install aiacc_nccl-2.0.0.tar.gz
# notes: must download and then pip install, cannot merge in oneline `pip install aiacc_xxx_url` 
# Both method1 and method2 can run concurrently.

• method3: python-pypi

pip install aiacc_nccl==2.0

Usage

After install aiacc-nccl package, you need do nothing to change code!

Environment

• AIACC_FASTTUNING: Enable Fasttuning for LLMs, default=1 is to enable.
• NCCL_AIACC_ALLREDUCE_DISABLE: Disable allreduce algo, default=0 is to enable.
• NCCL_AIACC_ALLGATHER_DISABLE: Disable allgather algo, default=0 is to enable.
• NCCL_AIACC_REDUCE_SCATTER_DISABLE: Disable reduce_scatter algo, default=0 is to enable.
• AIACC_UPDATE_ALGO_DISABLE: Disable update aiacc nccl algo from aiacc-sql-server, default=0 is to enable.

Performance

AIACC-NCCL can speedup the nccl performance on aliyun EGS(GPU machine), for example instance type 'ecs.ebmgn7ex.32xlarge' is A100 x 8 GPU and using network eRdma.

GPU(EGS)	Collective	Nodes	Network	Speedup(nccl-tests)
A100 x 8	all_gather	2-10	VPC/eRdma	30%+
A100 x 8	reduce_scatter	2-10	VPC/eRdma	30%+
A100 x 8	all_reduce	2-10	VPC/eRdma	20%
V100 x 8	all_reduce	2-20	VPC	60%+
A10 x 8	all_reduce	1	-	20%

Copyright

All source code and accompanying documentation is copyright (c) 2015-2020, NVIDIA CORPORATION. All rights reserved. All modifications are copyright (c) 2020-2024, ALIYUN CORPORATION. All rights reserved.