flink-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Stephan Ewen <se...@apache.org>
Subject Re: Stale Synchronous Parallel iterations in Flink
Date Mon, 23 Feb 2015 09:32:11 GMT
Hey Tran Nam-Luc!

Great post with some really cool thoughts.
I just posted this answer to your LinkedIN post.



Nice post, very cool idea! Your understanding of Flink in that respect is
really good. I had not heard of SSP before,
but it seems to be a good compromise between bulk synchronous and
asynchronous iterations.

Here are some comments and thought about how Apache Flink realizes
iterations and how that mechanism can be extended
to support SSP:

 - The loop in Flink is standing, operators are not re-created and
re-deployed in every iteration.

 - We signal the end of a superstep by pushing a special message from the
"head" of the loop to the "tail". This can be thought of
   like a clock cycle. Currently, the new superstep starts on each parallel
thread once all "tails" have received the message,
   thus forming the BSP barrier. For delta iterations, this "tail" is the
next workset - the solution set completely independent of that.

 - We can probably interpret the "end-of-superstep" messages as clock
messages. We could then allow threads to start their
   next superstep if all "tails" have seen clock messages at least of its
own clock time minus the slack.

If you are looking to implement this in Flink, or dig deeper into this, let
me know, I would be happy to help.

On Fri, Feb 20, 2015 at 5:27 PM, Nam-Luc Tran <namluc.tran@euranova.eu>

> Hello Everyone,
> I am Nam-Luc Tran, research Engineer at EURA NOVA [1]. Our research
> subjects cover distributed machine learning and we have been working
> on dataflow graph processing for a while now. We have been reading
> from you since Stratosphere :-)
> Our current research focuses on Stale Synchronous Parallelism and we
> are currently considering Apache Flink as a good candidate for
> implementing and delivering the best results among the existing
> processing solutions. I have written a post about it here:
> https://www.linkedin.com/pulse/stale-synchronous-parallelism-new-frontier-apache-flink-nam-luc-tran
> What do you guys think about the approach? Does it seem feasible, or
> do you have anything similar in your roadmap?
> Best regards,
> Tran Nam-Luc
> Links:
> ------
> [1] http://euranova.eu

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