AWS AWSEC2 documentation change
Summary
Updated EFA changelog entries with detailed release notes for libnccl-ofi, libfabric, efa driver, and rdma-core upgrades. Added performance optimizations, platform support details, and deprecation notices.
Security assessment
Changes focus on performance improvements, compatibility updates, and feature enhancements. While some entries mention fixes like 'Fix work request index double use', there's no explicit mention of security vulnerabilities or patches. Updates appear routine without security-specific context.
Diff
diff --git a/AWSEC2/latest/UserGuide/efa-changelog.md b/AWSEC2/latest/UserGuide/efa-changelog.md index 5eeeb458a..547bec30b 100644 --- a//AWSEC2/latest/UserGuide/efa-changelog.md +++ b//AWSEC2/latest/UserGuide/efa-changelog.md @@ -13 +13,4 @@ Version | Changes | Release date - * Upgrade to libnccl-ofi 1.16.3 + * **Upgrade to[ libnccl-ofi 1.16.3](https://github.com/aws/aws-ofi-nccl/releases/tag/v1.16.3)** + * Supports NCCL v2.27.7-1 while maintaining backward compatibility with NCCL v2.17.1 and later. + * Building with platform-aws requires Libfabric v1.22.0amzn4.0 or greater. Tested with versions up to Libfabric 2.1.0amzn5.0. + * Enable domain-per-thread by default on all Amazon EC2 instance types for improved performance for some applications where NCCL creates multiple proxy threads @@ -18,2 +21,5 @@ Version | Changes | Release date - * Upgrade to libfabric 2.1.0amzn5.0 - * Upgrade to libnccl-ofi 1.16.2 + * **Upgrade to[ libnccl-ofi 1.16.2](https://github.com/aws/aws-ofi-nccl/releases/tag/v1.16.2)** + * Supports NCCL v2.27.6-1 while maintaining backward compatibility with NCCL v2.17.1 and later. + * Building with platform-aws requires Libfabric v1.22.0amzn4.0 or greater. Tested with versions up to Libfabric 2.1.0amzn4.0. + * Added a new platform configuration to support using the OFI NCCL plugin on the Amazon EC2 p5.4xlarge instance type + * **Upgrade to[ libfabric 2.1.0amzn5.0](https://github.com/aws/libfabric/releases/tag/v2.1.0amzn5.0)** @@ -24,7 +30,13 @@ Version | Changes | Release date - * Deprecate support for Ubuntu 20.04 - * Upgrade to libnccl-ofi 1.16.1 - * Upgrade to rdma-core 58.amzn0 - * Upgrade to libfabric 2.1.0amzn4.0 - * Enable optimizations for Graviton platforms - * Upgrade to efa driver 2.17.2 - * Upgrade to efa-nv-peermem 1.2.2 + * **Upgrade to[ libnccl-ofi 1.16.1](https://github.com/aws/aws-ofi-nccl/releases/tag/v1.16.1)** + * Supports NCCL v2.27.6-1 while maintaining backward compatibility with NCCL v2.17.1 and later. + * Building with platform-aws requires Libfabric v1.22.0amzn4.0 or greater. Tested with versions up to Libfabric 2.1.0amzn3. + * Update the PCI link speed format reported in the topology file to match kernel 5.7+ + * Added SKIP_NICS_WITHOUT_ACCEL_AT_SAME_PCI_LEVEL to skip libfabric nics that do not have an accelerator at the same pci level + * **Upgrade to[ efa driver 2.17.2](https://github.com/amzn/amzn-drivers/releases/tag/efa_linux_2.17.2)** + * **Upgrade to[ efa-nv-peermem 1.2.2](https://github.com/amzn/amzn-drivers/releases/tag/efa_nv_peermem_linux_1.2.2)** + * **Upgrade to[ libfabric 2.1.0amzn4.0](https://github.com/aws/libfabric/releases/tag/v2.1.0amzn4.0)** + * **Upgrade to[ rdma-core 58.amzn0](https://github.com/linux-rdma/rdma-core/releases/tag/v58.0)** + * Fix work request index double use + * Add WQE length to post_send tracepoint + * **Enable optimizations for Graviton platforms** + * **Deprecate support for Ubuntu 20.04** @@ -35,5 +47,17 @@ Version | Changes | Release date - * Upgrade to efa driver 2.15.3 - * Upgrade to efa-nv-peermem 1.2.1 - * Upgrade to rdma-core 57.amzn1 - * Upgrade to libfabric 2.1.0amzn3.0 - * Upgrade to libnccl-ofi 1.15.0 + * **Upgrade to[ efa driver 2.15.3](https://github.com/amzn/amzn-drivers/releases/tag/efa_linux_2.15.3)** + * **Upgrade to[ efa-nv-peermem 1.2.1](https://github.com/amzn/amzn-drivers/releases/tag/efa_nv_peermem_linux_1.2.1)** + * **Upgrade to[ rdma-core 57.amzn1](https://github.com/linux-rdma/rdma-core/releases/tag/v57.1)** + * Fix work request index double use + * **Upgrade to[ libfabric 2.1.0amzn3.0](https://github.com/aws/libfabric/releases/tag/v2.1.0amzn3.0)** + * **Upgrade to[ libnccl-ofi 1.15.0](https://github.com/aws/aws-ofi-nccl/releases/tag/v1.15.0)** + * Supports NCCL 2.26.6-1 while maintaining backward compatibility with (NCCL v2.17.1 and later. + * Building with platform-aws requires Libfabric v1.22.0amzn4.0 or greater. Tested with versions up to Libfabric 2.1.0amzn3. + * Build system and platform support + * Added Amazon EC2 P6-B200 platform support + * Changed default plugin library name to libnccl-net-ofi.so, and by default create symlink from libnccl-net-ofi.so to libnccl-net.so to maintain backward compatibility. This allows users to set NCCL_NET_PLUGIN=ofi to force NCCL to use the OFI plugin for communication. Specifying --disable-nccl-net-symlink to configure will skip the symlink, allowing multiple plugins to be installed in the same container. + * Tuning and performance improvements + * Added tuner support on P6-B200 for AllReduce, AllGather, and ReduceScatter regions for 0x0 and 0x7 bitmask + * Updated default latency for P5en and P6-B200 platforms based on empirical results and analysis + * Update to use NCCL v10 API with trafficClass parameter support for future traffic prioritization + * Migrated plugin code base from C to C++ + * Added support for jobs where the number of NICs per GPU is different across systems. See the OFI_NCCL_FORCE_NUM_RAILS runtime environment variable documentation for more information.