Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Mon, 19 Dec 2011 23:36:40 +0400
From: Solar Designer <>
Subject: Re: DES BS + OMP improvements, MPI direction

On Mon, Dec 19, 2011 at 08:48:34AM -0700, RB wrote:
> For posterity's completeness, I meant that for a given run of X
> seconds, no given step/pass/stage (let's call it a 'set' unless you
> desire otherwise) that takes a single processor longer than X seconds
> to solve will finish, no matter the number of processor resources.
> This can be exacerbated in MPI environments with asymmetric processors
> (or even cooling, in the case of clock-boosting processors).
> It is often the case with MPI runs that, when aborted, many (if not
> all) of the processors have been working for a considerable length of
> time on a single set that, given the overall processing power
> available, "could" have been completed had all resources been focused
> there.

It is unclear from what you wrote above whether having completed that
single set would be any better or not.  You seem to imply that it would,
but this is not so obvious.

That said, overall I agree that the current MPI parallelization of
incremental mode is far from optimal when the number of nodes is large.

> It would be fascinating (but programmatically challenging) to
> have the implementation automatically switch approaches (or reduce
> working unit size) as timing for individual sets clears a given
> threshold.

I previously proposed how to do something like that (and even better) in
the last paragraph of:

In fact, I had this implemented in a working prototype in 1997 (I just
took a look at those files - the timestamps are Feb-Apr 1997), but I
never put any further time into it.


Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.