just carrying on from where I left off (I'm starting to write more than even Nflight usually does! ;) ...

If you are after an environment that would crunch fast and report back to the host then you could just have a load of loosely coupled (on a LAN or Internet) fast servers and a job scheduler that drops the tasks onto them and then waits for the results. Typically called High Throughput Computing (sort of similar to Distributed Computing - like BOINC then!), you would use this as an environment which handles many independent coarse tasks which get distributed in serial, the design would benefit from bandwidth due to the direct affect of high throughput (large tasks) and mass (lots of servers) but latency is not that important as each task is independent.

If you really want an environment which is more of a high computation environment then parallelised computing grids or HPC (high performance computing) would be the way forward (start saving as this is not cheap) these are designed to handle fine-grained tasks and have rapid synchronisation of the bits(tasks) due to running them in parallel. These are tightly coupled environments and would not suit a BOINC environment unless the developer designed the task specifically for it. Latency is the killer here, you need very low latency as a task is spread in pieces across so many processors that syncronisation of results is paramount and so the bottleneck that typically dictates the overall performance (not the compute power) is the connectivity. Inifiband is popular in these environments due to being fairly low latency and decent bandwidth (although there are a few new technologies coming out in 2007 and 2008 which would stomp all over Infiniband), I wouldn't even use 10G Ethernet as although you would get fairly decent bandwidth you would be hindered by the packet verification and addressing latency of Ethernet.


So, what would a scientific client want? depends how they constructed their application - either fine (typically parallel) or coarse grained (typically serial) tasks!