Bookshelf

In today’s data-driven world, high-performance computing has become more important than ever. Distributing work is the most successful strategy to process the huge amounts of data that need to be analzyed. As a response, modern processors comprise an increasing number of cores to be suited for parallel computation. The many cores require high throughput to and from the memory. However, the classical UMA architecture does not fit the requirement of higher bandwidths. This is because the path to the memory is uniform for all cores, which is why shorter and distributed paths to memory are prohibited. The NUMA archticture is a recent solution to mitigate this bottleneck by breaking with the uniformity of the memory accesses. The distinction between fast, local accesses and global but slower ones allows for higher bandwidths.

To benefit from the transition to NUMA, engineers have to rethink libraries, programs, and operating systems. With NUMA, some long-standing optimization models are invalidated and need to be revised. This thesis aims at guiding engineers who develop software for this new platform. In the form of a survey, a relevant amount of related work is characterized and classified. From the essence of the survey, general guidelines for optimizing software on NUMA systems are devised. As a practical example, this thesis provides a brief demonstration of how to apply these guidelines in a typical use case.