[Swift-user] When to use Coaster filesystem data provider?

Mihael Hategan hategan at mcs.anl.gov
Thu Jan 29 12:17:18 CST 2009


On Thu, 2009-01-29 at 11:47 -0600, Michael Wilde wrote:
> Please explain (on this list and/or in the userguide) more about this 
> data provider method:
> 
> - what are its performance characteristics?

Similar to coaster execution in that authentication cost is amortized
and high parallelism (say, >128) can be achieved. This means that for
small files you can get near 100% bandwidth utilization.

> 
> - when would one select it vs gridftp?

When there are many small files involved and you are willing to deal
with not-so-well-tested software.

> 
> - how does it work?

File data is sent as messages through the same communication library
that coasters use. In principle there is one TCP/SSL connection used for
two-way tagged messages, so RTT is amortized with high parallelism.

> 
> Was this implemented recently, or has it been in the coaster 
> implementation for quite a while?

It's been there since I was trying to speed up skenny's 64k jobs
workflow on ranger.
See
http://mail.ci.uchicago.edu/mailman/private/swift-devel/2008-September/003879.html





More information about the Swift-user mailing list