40555 Rev. 3.00 June 2006 | Performance Guidelines for AMD Athlon™ 64 and AMD Opteron™ |
| ccNUMA Multiprocessor Systems |
Chapter 1 Introduction
The AMD Athlon™ 64 and AMD Opteron™ family of
In traditional symmetric multiprocessing (SMP) systems, the various processors share a single memory controller. This single memory connection can become a performance bottleneck when all processors access memory at once. At the same time, the SMP architecture does not scale well into larger systems with a greater number of processors. The AMD ccNUMA architecture is designed to overcome these inherent SMP performance bottlenecks. It is a mature architecture that is designed to extract greater performance potential from multiprocessor systems.
As developers deploy more demanding workloads on these multiprocessor systems, common performance questions arise: Where should threads or processes be scheduled (thread or process placement)? Where should memory be allocated (memory placement)? The underlying operating system (OS), tuned for AMD Athlon 64 and AMD Opteron multiprocessor ccNUMA systems, makes these performance decisions transparent and easy.
Advanced developers, however, should be aware of the more advanced tools and techniques available for performance tuning. In addition to recommending mechanisms provided by the OS for explicit thread (or process) and memory placement, this application note explores advanced techniques such as node interleaving of memory to boost performance. This document also delves into the characterization of an AMD ccNUMA multiprocessor system, providing advanced developers with an understanding of the fundamentals necessary to enhance the performance of synthetic and real applications and to develop advanced tools.
In general, applications can be memory latency sensitive or memory bandwidth sensitive; both classes are important for performance tuning. In a multiprocessor system, in addition to memory latency and memory bandwidth, other factors influence performance:
•the latency of remote memory access (hop latency)
•the latency of maintaining cache coherence (probe latency)
•the bandwidth of the HyperTransport interconnect links
•the lengths of various buffer queues in the system
The empirical analysis presented in this document is based upon data provided by running a multi- threaded synthetic test. While this test is neither a pure memory latency test nor a pure memory
Chapter 1 | Introduction | 9 |