gapv64@udcf.glasgow.ac.uk ("B.Ewins") (02/14/91)
Sitting in front of me as I write is a paper from the Journal Of Parallel And Distributed Computing, Vol 11,No 1,(ie Jan 1991) entitled 'Communications Overhead and the Expected Speedup of Multidimensional Mesh-Connected Parallel Processors'. I can't say I'm impressed. What this paper does is extends Amdahl's law (That the parallel part of your program runs N times faster on N parallel processors) to include delays caused by having to talk to other processors, espescially those for 'particle problems' where distant processors must communicate. No offence to messrs Scherson+Corbett, but this isn't very advanced stuff. I don't mean that someone has done it before, just that I'm surprised no-one has. Now, I hadn't learnt what CSP was until the recent conversation on the net, since, having inherited a large parallel C program to work on I never felt the need to read up on occam.However, this conversation described CSP as an 'elegant mathematical tool' for understanding parallel programming. What I would like to know is, has any work been done on the mathematics of , basically 'how big should I make my network' , in the light of the possible increased communications load,blocking,deadlock etc that go al along with increased nos of processors ? It's a real question- the program I work on has rapidly increasing communications time with network size, hence, since the computational time decreases, there is an optimum no. of transputers. It's easy for me to find this since resources have not been a problem: but for others buying large systems may be a big mistake. Anyway, I've asked my stupid question, I expect an avalanche of informed answers. Brian Ewins,Glasgow Nuclear Structure Group. 'Whose matrices know no natural form, Unless subjected to prodigious strain:'-D.Davie