Crisis in HPC Discussion - Cliff Addison, Institute of Advanced Scientific Computation

Newsgroups: uk.org.epsrc.hpc.discussion
From: cliff@liverpool.ac.uk (Dr C. Addison)
Subject: Re: Crisis-in-HPC conclusions
Organization: The University of Liverpool
Message-ID: 

In article 35, lyndon@epcc.ed.ac.uk (L J Clarke) writes:
> * Efficiency of T3D - it's my experience that the performance of T3D programs
>   is determined by the efficiency of the single processor performance. People
>   are reporting efficiencies in the region of 17%, and this is more or less
>   just the single node efficiencies that codes are seeing. Its really got
>   nothing to do with MPP aspects of T3D, its to do with memory hierarchies.
>

This is a point that needs to be repeated and repeated. Sparse matrix problems, for instance solving linear systems via conjugate gradients, can completely thrash a lot of superscalar superpipelined processors because there is a lot of indirect addressing that can cause pipeline stalls and inefficient cache utilisation (1 word out of a cache line used etc. etc.).

A great deal of attention has been paid by developers of numerical software to exploit locality of reference and to get reasonable levels of performance >from a range of processors in an easily configured manner. A lot of this work involves fairly significant changes to the underlying algorithms and data structures employed. Excellent progress has been made in some problem areas, such as solving dense systems of linear equations, and a good deal of new generation software with related new algorithms are appearing, but there are some big gaps in our ability to exploit a single processor efficiently.

I really think much more work needs to be supported in this area, with an secondary objective of designing these new algorithms so they can cope with several levels of memory hierarchy and hence have the potential to provide a reasonable basis for good parallel software. Developing parallel software and tools without addressing this underlying problem is building on a foundation of sand -- things are liable to slip from underneath you.

Cliff Addison
Institute of Advanced Scientific Computation
Victoria Building, Brownlow Hill
University of Liverpool
Liverpool L69 3BX
U.K.

E-mail: cliff@liverpool.ac.uk
Tel: +44 151 794 4738
FAX: +44 151 794 4754