On Mon, Dec 08, 2008 at 01:44:05PM +0000, John Bland wrote:
> A quick note that following Kostas' advice we tried a very small buffer
> size, 4kB. This reduced bandwidth consumption even further but cpu
> efficiency was still relatively poor (hitting network latency rather
> than local RAM). Our advice is still to try the big buffer (128MB)
> unless your network is significantly bottle necking.
Does the real time of the job stay the same for all buffer sizes? Just
because it's more efficient it doesn't mean that it finishes sooner. A
lot of cpu could be consumed in the buffer management, I have no idea if
dcap/rfio uses an LRU or another less/more advanced algotithm but I'll
surprised if they are optimised for buffers that big and a tiny hit
PS> Read http://northgrid-tech.blogspot.com/2008/12/rfio-tuning-for-atlas-analysis-jobs.html
for the tests/results.