[Darshan-users] Q about darshan-analyzer and ranks
Shane Snyder
ssnyder at mcs.anl.gov
Fri Jul 8 12:49:58 CDT 2016
Hi Caz,
Not silly at all, that is a little confusing :)
In 2.x versions of Darshan, 'multi' would indicate that a given Darshan
log contained file records that were not opened collectively (across all
ranks, denoted by rank == -1). 'single' would indicate that a log
contained file records that were opened collectively across all ranks. I
*think* the thought process for naming that way is that files which are
accessed by all ranks have a *single* record, whereas applications which
open independent files on each process have *multiple* file records.
At any rate, in newer versions of Darshan (>=3.0.0), the wording has
been changed to be more clear. We use 'shared' to denote applications
which used a single, shared file for I/O (i.e., single=>shared) and we
use 'fpp' (file-per-process) to indicate logs which have files that were
opened independently by a certain process (i.e., multi=>fpp).
Also, note that a given log file can be viewed as both single/shared and
multi/fpp by darshan-analyzer. Any log that has a shared file record in
it will be classified as single/shared and similarly any log with a file
that was accessed by either a single rank or a subset of ranks will be
classified as multi/fpp. The tool is just meant to give a general
overview of the type of access patterns a directory of logs used.
--Shane
On 07/08/2016 11:47 AM, Casmira, Jason P wrote:
>
> Hello.
>
> Hopefully this isn’t too silly of a question.
>
> I see this code in darshan-analyzer.c:
>
> if (cp_file.rank == -1)
>
> *used_single = 1;
>
> else
>
> *used_multi = 1;
>
> But from darshan_parser.c:
>
> printf("\n# description of columns:\n");
>
> printf("# <rank>: MPI rank. -1 indicates that the file is
> shared\n");
>
> printf("# across all processes and statistics are
> aggregated.\n");
>
> It’s not clear to me how to interpret darshan_analyzer’s use of the
> rank field.
>
> Intuitively I’m thinking it would be the opposite (-1 == shared ==
> multi), so I think
>
> I’m missing something fundamental here, and I hope someone can clarify
> for me.
>
> I tried searching on “rank” and “darshan-analyzer” but couldn’t find
> an answer.
>
> Thanks.
>
> Caz.
>
> Jason Casmira
>
> Intel Corporation
>
> 503.712.5610
>
>
>
> _______________________________________________
> Darshan-users mailing list
> Darshan-users at lists.mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/darshan-users
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/darshan-users/attachments/20160708/5545ecae/attachment.html>
More information about the Darshan-users
mailing list