[Fwd: Re: [Swift-devel] Re: swift-falkon problem... plots to explain plateaus...]

[Ben Clifford]

On Tue, 1 Apr 2008, Mihael Hategan wrote:

> Having optimized wrappers for different architectures is a perfectly
> valid option.

it might also be possible to replace the entire execute2 layer of 
stagein-execute-stageout if falkon wants to do its own worker-node data 
placement - Swift would be the same down to calling execute2 but the 
execute2-replacement would be falkon specific rather than using the 
present model which assumes a shared filesystem for stageins.

I think that might fit in better with what Falkon is trying to do, letting 
it know which files are required by which job, rather than assuming a 
cluster-wide shared filesystem to fetch data from.

(The same might apply for using condor with no shared filesystem, which 
isthe situation in many campus workstation labs that I've seen - an 
execute2 layer that submits a bundled up stagein/stageout/execute as a 
single condor submission - I have been mulling over that for a few months 
and maybe will get someone to play with that in google summer of code)

--