[Beowulf] Good demo applications for small, slow cluster

Douglas Eadline deadline at eadline.org
Wed Aug 21 09:45:19 PDT 2013


> Sorts in general.. Good idea.
>
> Yes, we'll do a distributed computing bubble sort.
>
> Interesting, though.. There are probably simple algorithms which are
> efficient in a single processor environment, but become egregiously
> inefficient when distributed.

e.g. The NAS parallel suite has an integer sort (IS) that is
very latency sensitive.

For demo purposes, nothing beats parallel rendering.
There used to be a PVM and MPI POVRay packages
that demonstrated faster completion times as more nodes were
added.

--
Doug


>
> Jim
>
>
>
> On 8/20/13 12:11 PM, "Max R. Dechantsreiter" <max at performancejones.com>
> wrote:
>
>>Hi Jim,
>>
>>How about bucket sort?
>>
>>Make N as small as need be for cluster capability.
>>
>>Regards,
>>
>>Max
>>---
>>
>>
>>
>>On Tue, 20 Aug 2013 maxd at performancejones.com wrote:
>>
>>> Date: Tue, 20 Aug 2013 00:23:53 +0000
>>> From: "Lux, Jim (337C)" <james.p.lux at jpl.nasa.gov>
>>> Subject: [Beowulf] Good demo applications for small, slow cluster
>>> To: "beowulf at beowulf.org" <beowulf at beowulf.org>
>>> Message-ID:
>>> 	<F7B6AE13222F1B43B210AA4991836895236966E9 at ap-embx-sp40.RES.AD.JPL>
>>> Content-Type: text/plain; charset="us-ascii"
>>>
>>> I'm looking for some simple demo applications for a small, very slow
>>>cluster that would provide a good introduction to using message passing
>>>to implement parallelism.
>>>
>>> The processors are quite limited in performance (maybe a  few MFLOP),
>>>and they can be arranged in a variety of topologies (shared bus, rings,
>>>hypercube) with 3 network interfaces on each node.  The processor to
>>>processor link probably runs at about 1 Mbit/second, so sending 1 kByte
>>>takes 8 milliseconds
>>>
>>>
>>> So I'd like some computational problems that can be given as
>>>assignments on this toy cluster, and someone can thrash through getting
>>>it to work, and in the course of things, understand about things like
>>>bus contention, multihop vs single hop paths, distributing data and
>>>collecting results, etc.
>>>
>>> There's things like N-body gravity simulations, parallelized FFTs, and
>>>so forth.  All of these would run faster in parallel than serially on
>>>one node, and the performance should be strongly affected by the
>>>interconnect topology.  They also have real-world uses (so, while toys,
>>>they are representative of what people really do with clusters)
>>>
>>> Since sending data takes milliseconds, it seems that computational
>>>chunks which also take milliseconds is of the right scale.  And, of
>>>course, we could always slow down the communication, to look at the
>>>effect.
>>>
>>> There's no I/O on the nodes other than some LEDs, which could blink in
>>>different colors to indicate what's going on in that node (e.g.
>>>communicating, computing, waiting)
>>>
>>> Yes, this could all be done in simulation with virtual machines (and
>>>probably cheaper), but it's more visceral and tactile if you're
>>>physically connecting and disconnecting cables between nodes, and it's
>>>learning about error behaviors and such that's what I'm getting at.
>>>
>>> Kind of like doing biology dissection, physics lab or chem lab for
>>>real, as opposed to simulation.  You want the experience of "oops, I
>>>connected the cables in the wrong order"
>>>
>>> Jim Lux
>>>
>>_______________________________________________
>>Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
>>To change your subscription (digest mode or unsubscribe) visit
>>http://www.beowulf.org/mailman/listinfo/beowulf
>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
> To change your subscription (digest mode or unsubscribe) visit
> http://www.beowulf.org/mailman/listinfo/beowulf
>
> --
> Mailscanner: Clean
>


-- 
Doug

-- 
Mailscanner: Clean



More information about the Beowulf mailing list