Crisis in HPC Personal Comment - Tom Lake, GLOSSA & InterGlossa Ltd.
Thanks for organising the `Crisis in HPC' meeting - it was very
enjoyable. I think that there are some serious points coming out
of it.
Large machine and small machines
The T3D can achieve top power with a relatively small number of
processors using leading edge processors and interconnect. This
means that problems don't have to be TOO large to use it well.
Machines using less tricky technology might be more
cost-effective but would would need relatively larger problems to
be used well because they had more processors to achieve a given
power.
Funding committees might consider buying more cost-effective,
better quality (low g,l) machines at lower single node power.
BSP
I was suprised to hear everyone talking about the shared memory
library and little use of the BSP ideas on the T3D. Can this
really be the situation.
Your dream machine
There are a lot of intermediate stages between where we are today
and the full virtual shared memory story as I am sure you are aware.
I think it won't be possible to throw away the advantages of scheduling
the thread where the private data is already loaded for a long time.
How does your machine actually deal with cache coherence of channels?
Is this a special protocol? It would be harder in C with channels
because of the uncertainty over type.
Regards,
Tom Lake
-------------------------------------------------------------------------------
Parallel Processing and Research Consultants | GLOSSA & InterGlossa Ltd
Email: Tom.Lake@glossa.co.uk | 59, Alexandra Road
Fax: +44 1734 561919 | Reading RG1 5PG
Tel: +44 1734 561919 | UK
-------------------------------------------------------------------------------