[codes-ross-users] Too many rollback events

[Jian Peng]

Thank you Nikhil, the nkp option works pretty well!

On Mon, Jun 5, 2017 at 10:13 AM, Nikhil Jain <nikhil.jain at acm.org> wrote:

> I don’t think 10 microsecond is a large lookahead; How big a system are
> you trying to simulate (as in the network size)? If the system is small, it
> may be hard to reduce the rollbacks as the #core increase.
>
> Other things that can be tried:
> 1. nkp option = number of KPs; but default ROSS uses 16, but you can set
> it to (#total LPs)/(#cores) - i.e. make each LP has its own KP, which will
> prevent false rollbacks
> 2. gvt/batch size = you can try a small gvt frequency (like 4/8) and batch
> size (4/8).
>
> ---
> Nikhil Jain
> Postdoctoral Fellow, Lawrence Livermore National Laboratory
> nikhil.jain at acm.org, http://nikhil-jain.github.io/
>
> On Jun 5, 2017, at 11:57, Jian Peng <jpeng10 at hawk.iit.edu> wrote:
>
> Hi,
>
>     My simulation is having too many events rolled back, like 70%. It's
> still relatively small scale, 16 cores on 4 nodes, using optimistic
> scheduler.  The result is attached.
>
>     Any general idea of reducing number of rollback events? I used the
> max-opt-lookahead option and set it to 10000. Is it still a large number?
>
>     Thanks!
> <rollback_result_1.png><rollback_result_2.png>________
> _______________________________________
> codes-ross-users mailing list
> codes-ross-users at lists.mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/codes-ross-users
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/codes-ross-users/attachments/20170607/9ed20d81/attachment.html>