[mpich2-dev] modify mpich2 aggregator

[Jaln]

Hi Dr. Latham,
Thanks for your suggestion. I have another question regarding the second
phase for collective read. After timing each process's cost in the shuffle
phase, I found that the aggregator usually finishes first, then
non-aggregators finishes later. Is it because the aggregator first send the
data to a buffer then other processes receive the data from the buffer? If
it's true, can I say, each aggregator's shuffle cost equals the aggregator'
sending cost plus the maximum receiving cost of other processes?

Thanks,

Jialin


On Thu, Oct 25, 2012 at 12:47 PM, Rob Latham <robl at mcs.anl.gov> wrote:

> On Thu, Oct 25, 2012 at 12:05:00PM -0500, Jaln wrote:
> > Hello everybody,
> >
> > I want to evaluate some ideas regarding the I/O aggregators. But have no
> > idea how to modify the mpich2 and pvfs2.
> > The idea basically is to pass some hint from mpich to pvfs layer.
> > What I want to do are:
> > 1. For collective read, e.g., in 'ad_read_coll.c', I can see the
> two-phase
> > I/O is implemented there. I want to assign some values for each
> aggregator,
> > e.g., priority value for various aggregators.
> > 2. Still for read, I want the assigned value can be seen by the
> underlying
> > file systems' job scheduler. e.g., in pvfs2, src/io/job/job.c,
> > then the file system's job scheduler can schedule the aggregators by the
> > assigned value.
>
> PVFS2 has a facility for end-to-end tracing, but it's not often used.
> I think the pvfs2-developers mailing list might be a good place to
> discuss that feature.
>
> You will probably have to modify ad_read_coll to make explicit
> PVFS_sys_io calls.
>
> ==rob
>
>
> --
> Rob Latham
> Mathematics and Computer Science Division
> Argonne National Lab, IL USA
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/mpich2-dev/attachments/20121106/573ed693/attachment.html>