Open cnettel opened 8 years ago
It would make sense and give more accurate results if we would have a global history instead of per-worker, but who would keep that global history in memory, the master? For large histories this could be a problem for performance, I guess. But yeah, if performance is not an issue, I think we should change to global hitrate history.
On Sun, Apr 3, 2016 at 7:49 AM, cnettel notifications@github.com wrote:
The hit rate history is per worker. This means that the meaning of hitrate2000 varies with the number of nodes. Another, complicating, issue is that events are not round-robined between workers in a fair way (with the psana backend). Thus, some hit history is trailing a long time.
Should the history be changed to be a total, not per worker, since the hitrate code anyway does an MPI reduction? This would give more comparable results when running in single or MPI mode, or varying the MPI width.
b
The hit rate history is per worker. This means that the meaning of hitrate2000 varies with the number of nodes. Another, complicating, issue is that events are not round-robined between workers in a fair way (with the psana backend). Thus, some hit history is trailing a long time.
Should the history be changed to be a total, not per worker, since the hitrate code anyway does an MPI reduction? This would give more comparable results when running in single or MPI mode, or varying the MPI width.