NVIDIA Unveils NCCL 2.22 with Enhanced Memory Efficiency and Faster Initialization

The NVIDIA Collective Communications Library (NCCL) has released its latest version, NCCL 2.22, bringing significant enhancements aimed at optimizing memory usage, accelerating initialization times, and introducing a cost estimation API. These updates are crucial for high-performance computing (HPC) and artificial intelligence (AI) applications, according to the NVIDIA Technical Blog.

Release Highlights

NVIDIA Magnum IO NCCL is designed to optimize inter-GPU and multi-node communication, which is essential for efficient parallel computing. Key features of the NCCL 2.22 release include:

• Lazy Connection Establishment: This feature delays the creation of connections until they are needed, significantly reducing GPU memory overhead.
• New API for Cost Estimation: A new API helps optimize compute and communication overlap or research the NCCL cost model.
• Optimizations for ncclCommInitRank: Redundant topology queries are eliminated, speeding up initialization by up to 90% for applications creating multiple communicators.
• Support for Multiple Subnets with IB Router: Adds support for communication in jobs spanning multiple InfiniBand subnets, enabling larger DL training jobs.

Features in Detail

Lazy Connection Establishment

NCCL 2.22 introduces lazy connection establishment, which significantly reduces GPU memory usage by delaying the creation of connections until they are actually needed. This feature is particularly beneficial for applications that use a narrow scope, such as running the same algorithm repeatedly. The feature is enabled by default but can be disabled by setting NCCL_RUNTIME_CONNECT=0.

New Cost Model API

The new API, ncclGroupSimulateEnd, allows developers to estimate the time required for operations, aiding in the optimization of compute and communication overlap. While the estimates may not perfectly align with reality, they provide a useful guideline for performance tuning.

Initialization Optimizations

To minimize initialization overhead, the NCCL team has introduced several optimizations, including lazy connection establishment and intra-node topology fusion. These improvements can reduce ncclCommInitRank execution time by up to 90%, making it significantly faster for applications that create multiple communicators.

New Tuner Plugin Interface

The new tuner plugin interface (v3) provides a per-collective 2D cost table, reporting the estimated time needed for operations. This allows external tuners to optimize algorithm and protocol combinations for better performance.

Static Plugin Linking

For convenience and to avoid loading issues, NCCL 2.22 supports static linking of network or tuner plugins. Applications can specify this by setting NCCL_NET_PLUGIN or NCCL_TUNER_PLUGIN to STATIC_PLUGIN.

Group Semantics for Abort or Destroy

NCCL 2.22 introduces group semantics for ncclCommDestroy and ncclCommAbort, allowing multiple communicators to be destroyed simultaneously. This feature aims to prevent deadlocks and improve user experience.

IB Router Support

With this release, NCCL can operate across different InfiniBand subnets, enhancing communication for larger networks. The library automatically detects and establishes connections between endpoints on different subnets, using FLID for higher performance and adaptive routing.

Bug Fixes and Minor Updates

The NCCL 2.22 release also includes several bug fixes and minor updates:

• Support for the allreduce tree algorithm on DGX Google Cloud.
• Logging of NIC names in IB async errors.
• Improved performance of registered send and receive operations.
• Added infrastructure code for NVIDIA Trusted Computing Solutions.
• Separate traffic class for IB and RoCE control messages to enable advanced QoS.
• Support for PCI peer-to-peer communications across partitioned Broadcom PCI switches.

Summary

The NCCL 2.22 release introduces several significant features and optimizations aimed at improving performance and efficiency for HPC and AI applications. The improvements include a new tuner plugin interface, support for static linking of plugins, and enhanced group semantics to prevent deadlocks.