incubator-general mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ted Dunning <>
Subject Re: Re: [DISCUSS] [PROPOSAL] Singa for Apache Incubator
Date Wed, 11 Feb 2015 01:22:39 GMT
I didn't read the documentation carefully enough to quote details, but it
appeared that the efficiency and asynchronous nature of the parameter
server is considered to be a key factor in scalability and performance.
The performance numbers that I read compared singa to H2O and showed a very
considerable performance advantage for singa.  This is very interesting
because H2O is a very high performance system.

Another factor that stood out in my reading was the fact that while Singa
performs very, very well on the problems shown, scaling breaks down a bit
at moderate cluster sizes.  The commentary in the report attributes this to
limitations in parameter server scaling.  I find it an intriguing
hypothesis that the synchronous updates that H2O does may well cause much
the same effect as a poorly scaled parameter server.

On Tue, Feb 10, 2015 at 3:59 PM, Edward J. Yoon <>

> > I think that, one of the big differences is that Singa is written in C++.
> Awesome, I'd be the first client. And anything from architectural
> viewpoint?
> --
> Best Regards, Edward J. Yoon
> -----Original Message-----
> From: Ted Dunning []
> Sent: Wednesday, February 11, 2015 8:37 AM
> To:
> Subject: Re: Re: [DISCUSS] [PROPOSAL] Singa for Apache Incubator
> On Tue, Feb 10, 2015 at 3:33 PM, Edward J. Yoon <>
> wrote:
> > My coworker is working on implementing DNN on Apache Hama (which supports
> > general-purpose BSP computing and Pregel-like graph framework). If Hama
> is
> > leveraging InfiniBand and GPUs in the future, what will be the major
> > difference bt Hama-based DistBelief clone and Singa project?
> >
> > And, how much performance difference bt async and sync?
> >
> I think that, one of the big differences is that Singa is written in C++.
> Another is that Singa runs now.  The present tense is always very powerful
> when it comes to software.
> ---------------------------------------------------------------------
> To unsubscribe, e-mail:
> For additional commands, e-mail:

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