Ian Turton
Centre for Computational Geography
School of Geography
University of Leeds
Leeds LS2 9JT
UK
Email: ian@geog.leeds.ac.uk
Phone: +44 (0)113 2333309
Fax: +44 (0)113 2333308
URL: http://gam.leeds.ac.uk/staff/i.turton/i.turton.html
I am a researcher employed under the EPSRC High Performance Computing
Initiative. The work involves taking existing geographical codes and
porting them to the Cray T3D. We were initially awed by claims of
multi GigaFlop performance, however we soon became disillusioned by the
reality. Months of hard work hand optimising code to achieve 7 Gflops
on an embarrassingly parallel task on a machine rated at 38 Gflops.
Later codes proved even harder to attain even modest speeds of 5
GFlops. It appears the `natural' speed of an application is below 1
GFlop, and that a lot of work is required to move it above this.One problem that we have identified is the restriction in the data parallel mode of CRAFT that arrays be a power of two. This may be acceptable when the size of arrays is essentially arbitrary but for geography we can not simply create more zones as needed. There are 10,674 wards in Great Britain and forcing all the arrays in the application to be 16,384 leads to wasting one third of the machine. If the parallelism is applied wrongly then one third of the processors are idle, even when the array is laid out correctly approximately 20% of the processor time can be idle.
This leaves the difficult position that whilst even 5 or 7 Gflops is faster than these programs have ever gone before it isn't as fast as advertised and it was hard work to achieve them. At present it is very hard to see more geographers being encouraged to attempt to make use of HPC.