Compressive sampling

Project overview

Compressive sampling (also known as compressed sensing) is a nascent signal processing technique used to sample a signal that is sparse in some domain. The goals of our project in CS are the advancement of the science of CS and its application to technology concerning IBM products, particularly as they pertain to systems. Our scientific work focuses on finding and using maximally incoherent bases.

One application we focus on is CS for processor core performance gathering. We made the first proposal and examination of dynamic compression of information gathered from a performance monitor unit (PMU). Conventional compression mechanisms are impractical as they require first the acquisition of the entire signal in a processor core, and then the application of a possibly computationally intensive compression algorithm.

Practical compression requires little processing of the signal in the processor core, allowing the compression to be achieved in simple hardware logic. On the other hand, regular sampling of the signal, while simple, does not generally achieve high fidelity unless the rate of sampling is high. Performance information is highly structured, allowing it to be compressed, but the computational burden of conventional compression techniques exclude their use in this environment.

We evaluated the practicality of using such techniques in the transfer of signals representing one or more micro-architectural counters from a processor core and showed that CS is usable to recover such performance signals, evaluating the trade-off between efficiency, accuracy and practicability with its various variants. We identify deficiencies in the existing techniques and propose refinements, which may be used to enhance the use of compressive sampling for processor performance monitoring.

Selected publications

  1. Tomas Tuma, Sean Rooney, Paul Hurley
    “On the Applicability of Compressive Sampling in Fine Grained Processor Performance Monitoring,” 14th IEEE International Conference on Engineering of Complex Computer Systems, June 2009.
  2. Tomas Tuma, Paul Hurley
    “On the incoherence of noiselet and Haar bases,” 8th international conference on Sampling Theory and Applications, May 2009.