Compiler Technologies for High End Computing on Commodity CPUs

The growth of clusters in High End Computing has resulted in significant technological challenges. Increasing complexity in CPUs and memory systems increases the need for sophisticated compiler optimizations, while the rise of clusters has meant that many systems are now built out of building blocks that are "mixed and matched", not really designed to work well together. A prime example of this problem is systems built using the AMD Opteron. Although the Opteron has great performance, including great memory latency and bandwidth, AMD does not have an in-house compiler team, and existing 3rd party compilers for AMD's new 64-bit mode have weak performance. This talk will describe these issues in detail, and then describe PathScale's new compiler suite for the Opteron.