hama-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Suraj Menon <surajsme...@apache.org>
Subject Re: The Iteration and Superstep
Date Thu, 03 Jan 2013 15:35:16 GMT
>From what I found while refactoring the code for GraphJobRunner in HAMA-632
is that roughly (not looking at edge cases):

if (aggregationRunner.isEnabled) /// If we have aggregators setup the
values would be aggregated by master in the extra superstep.
iterations = 2 * (number of supersteps after all vertices are loaded,
partitioned and repaired)
else
iterations = (number of supersteps after all vertices are loaded,
partitioned and repaired)

I think we can maintain the last execution completion time in
Vertex.compute to find the time for each iteration. Shouldn't this work?

-Suraj


On Thu, Jan 3, 2013 at 8:34 AM, Shuo Wang <ecisp.wangshuo@gmail.com> wrote:

> But now we can't know the superstep is the *iteration* or the *
> additional superstep*.
>
> 2013/1/3 Edward J. Yoon <edwardyoon@apache.org>
>
> > > Do you mean that the time of each *interation* equals the time of per *
> > > superstep*?
> >
> > Yes, each iteration process is a superstep of BSP program.
> >
> > On Thu, Jan 3, 2013 at 10:21 PM, Shuo Wang <ecisp.wangshuo@gmail.com>
> > wrote:
> > > Do you mean that the time of each *interation* equals the time of per *
> > > superstep*?
> > >
> > > 2013/1/3 Edward J. Yoon <edwardyoon@apache.org>
> > >
> > >> We don't logging execution time of each iteration yet.
> > >>
> > >> But, it's not difficult. You can get current time before and after
> > >> calling doSuperstep() method, and then calculate execution time of
> > >> each iteration. If you want to improve our project, Pls feel free to
> > >> create JIRA tickets and attach your patches. ;)
> > >>
> > >> On Thu, Jan 3, 2013 at 9:52 PM, Shuo Wang <ecisp.wangshuo@gmail.com>
> > >> wrote:
> > >> > Year,I got it.But how can I kown the executed time per iteration?
> > >> >
> > >> > 2013/1/3 Edward J. Yoon <edwardyoon@apache.org>
> > >> >
> > >> >> (Pregel-like) graph computations consist of a sequence of
> > *iterations*.
> > >> >>
> > >> >> To perform the graph job, internally, we also do additional global
> > >> >> synchronizations. For example, counts and synchronizes the number
> of
> > >> >> total vertices among Workers.
> > >> >>
> > >> >> So, the total *supersteps* of BSP program for graph job =
> > *iterations*
> > >> >> + additional supersteps.
> > >> >>
> > >> >> On Thu, Jan 3, 2013 at 6:44 PM, Shuo Wang <
> ecisp.wangshuo@gmail.com>
> > >> >> wrote:
> > >> >> > Hi,
> > >> >> >
> > >> >> > What is the difference between the *iterations* and the
> > *supersteps*
> > >> when
> > >> >> > we run pagerank or sssp on HAMA?
> > >> >> >
> > >> >> > wangshuo
> > >> >>
> > >> >>
> > >> >>
> > >> >> --
> > >> >> Best Regards, Edward J. Yoon
> > >> >> @eddieyoon
> > >> >>
> > >>
> > >>
> > >>
> > >> --
> > >> Best Regards, Edward J. Yoon
> > >> @eddieyoon
> > >>
> >
> >
> >
> > --
> > Best Regards, Edward J. Yoon
> > @eddieyoon
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message