Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Fri, 29 Jan 2016 22:25:44 -0500
From: japhar81 <>
Subject: Re: MPI with Spot Instances?

They basically just disappear, whatever they were in the middle of -- i
guess my question is, will the resume re-run whatever jobs those nodes were
in the middle of and didn't report back? And if one of them happens to have
hit a match, will that get saved somehow too?

On Fri, Jan 29, 2016 at 10:20 PM, magnum <> wrote:

> On 2016-01-29 18:12, japhar81 wrote:
>> Ok, so corollary question -- does the session stuff work with MPI? i.e.
>> lets say I start the spot instances externally, and mpiexec jtr with some
>> flavor of --session (on a box that wont die). If those nodes die
>> mid-process, will that be recorded in the session to enable a resume later
>> when I spin new nodes and start mpiexec again?
> Sure (as far as I can imagine how spot instances work). Session file
> integrity is very well tested.
> magnum
> On Wed, Jan 27, 2016 at 4:03 PM, magnum <> wrote:
>> On 2016-01-27 17:25, japhar81 wrote:
>>> I've been playing around with MPI clustering JtR for a while, and I've
>>>> managed to get it running smoothly on static nodes. What I'd like to do
>>>> next is create an auto-scaling group in AWS, using spot instances. What
>>>> this basically means is nodes will come and go, with their hostnames/IPs
>>>> changing at random.. I can not figure out how I would run JtR in that
>>>> instance -- since it requires a node list in a file on startup to
>>>> mpirun.
>>>> If it matters, I'm looking to do a brute-force using the ASCII mode. Has
>>>> anyone found a way to do a dynamic cluster that adds/removes nodes at
>>>> random? Is this even possible?
>>> I'm not aware of any existing work that would do this. A solution using
>>> JtR as-is, but with some yet-to-be-implemented master issuing jobs, could
>>> involve looking at the existing "-node=x/y" as describing "pieces"
>>> instead
>>> of "nodes". So instead of saying -node=2/8 as in "you are node 2 of 8"
>>> you'd say -node=4321/100000 as in "do piece 4321 of 100000". Then you'd
>>> submit pieces to active nodes. Obviously you'd have to handle dying nodes
>>> that never reported back their given piece, and re-issue those pieces.
>>> magnum

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.