NUMA Optimizations

General Remarks

This project aims at developing programming and optimization strategies to improve data placement on shared memory computers with non-uniform data access characteristics. This includes ccNUMA systems, but also multi-core processors that often have a complex hierarchical cache structure.

C++ programming techniques for High Performance Computing on systems with non-uniform memory access using OpenMP

Diploma thesis at the Georg Simon Ohm University of Applied Sciences Nuremberg.

Author: Holger Stengel (now RRZE)

Supervisor: Georg Hager (HPC Services).

Abstract: This work develops programming methodologies for C++ that respect the need for optimal NUMA page placement in OpenMP code. An overloaded new[] operator is presented that guarantees proper placement for arrays of objects. Along the same lines, the STL vector<> class can be endowed with an allocator class argument that serves the same purpose. The disadvantages of std::vector<> in terms of performance and usability in a NUMA setting are circumvented by developing a special numa_vector<> container which is compatible with all STL algorithms. Finally, a container with a segmented, padded data structure, including appropriate iterators, allows one to make generic algorithms aware of data segmentation of any kind (including NUMA) without sacrificing performance.

Downloads: Poster (English), Thesis (in German).

Contact

Project manager:

• Dr. Georg Hager
  Regionales Rechenzentrum Erlangen
  HPC Services
  Martensstr. 1
  91058 Erlangen
  +49 (0)9131 85 28973
  georg.hager@rrze.fau.de