IntroductionNon-Uniform Memory Access (NUMA) is distributed shared memory architecture, where the access latency to memory is non-uniform across different processors. In NUMA, each processor or set of common cores will have its own local memory forming a small building block (called node/domain). These blocks are interconnected across high speed network to form a large scale Symmetric Multiprocessor (SMP) server. These nodes are called NUMA nodes or Resource Affinity Domains (RAD). OpenVMS support for NUMA was first introduced in version 7.2-1H1 for older Alpha systems and extended its support for Intel® Itanium® architecture based HP Integrity servers with the V8.4 release.
Refer to OpenVMS Technical Journal (VTJ) called RAD on OpenVMS for more details about OpenVMS support for NUMA systems.
This document provides detailed overview and usage of OpenVMS NUMA/RAD programming interfaces. The intent of this document is to show the usage of programming interfaces and ability to create custom Application Programming Interface (API). The techniques used in the sample APIs can help application designers to take maximum advantage of complex NUMA architectures. The application performance in NUMA environments largely depend on its architecture. The major impact happens when the application allocates resources which aren’t aligned with NUMA locality.
NUMA has become de facto architecture of all the latest scale-up SMP servers that are available in the market. Even most of the new-generation servers shipped by HP are NUMA based. Examples are the BL8x0c i2 blades and rx2800 i2 rack-mounted servers. Historically, HP is the market leader in manufacturing the NUMA based servers. A class of Alpha and some previous generation Itanium (IA64) servers were NUMA based. Examples of Alpha NUMA systems are GS series systems and examples of Intel Itanium architecture based HP Integrity servers are HP Integrity rx76xx/rx86xx series (also called as cell-based systems).
The OpenVMS system services allow the programmers to get the NUMA specific configuration and to provide NUMA hints to the system so that the resources are allocated aligning to the NUMA locality. The C source files mentioned in the Appendix rad_routines, rad_crmpsc, rad_creprc, and rad_qops are examples that demonstrate how to program with OpenVMS NUMA system service interfaces.
read more
More...