Beowulf and variable cpus
p.gri...@virgin.net
Hi, I am new to Beowulf and have some questions,
1. Does anyone know if I am able to run a beowulf cluster with
different types of clients, i.e I have a load of pentium 100s
and some p2 500s which I would like to use together. Is this
possible?
Regards,
Paul Grimshaw.
-----
Sent using MailStart.com ( http://MailStart.Com/welcome.html )
The FREE way to access your mailbox via any web browser, anywhere!
_______________________________________________
Beowulf mailing list
Beo...@beowulf.org
http://www.beowulf.org/mailman/listinfo/beowulf
Troy Baer
> Hi, I am new to Beowulf and have some questions,
>
> 1. Does anyone know if I am able to run a beowulf cluster with
> different types of clients, i.e I have a load of pentium 100s
> and some p2 500s which I would like to use together. Is this
> possible?
Sure, it's definitely possible. You may run into load balancing problems
if you have a parallel programs running on both fast and slow CPUs. If
you use a batch system like PBS, you can create attributes for individual
nodes and use those to ask a homogeneous set of processors (or
heterogeneous, for that matter):
# ask PBS for 10 Pentium 100MHz nodes
#PBS -l nodes=10:p_100
# ask PBS for 6 Pentium II 500 MHz nodes
#PBS -l nodes=6:p2_500
# ask PBS for 10 Pentium 100s *and* 6 P2 500s
#PBS -l nodes=10:p_100+6:p2_500
I'd assume/hope you can do similar things in other resource managers and
batch systems, like LSF and DQS (although I've never tried either).
--Troy
--
Troy Baer email: tr...@osc.edu
Science & Technology Support phone: 614-292-9701
Ohio Supercomputer Center web: http://oscinfo.osc.edu
Jag
--ZARJHfwaSJQLOEUz
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
Content-Transfer-Encoding: quoted-printable
On Thu, 21 Sep 2000, p.gri...@virgin.net wrote:
>=20
> Hi, I am new to Beowulf and have some questions,
>=20
> 1. Does anyone know if I am able to run a beowulf cluster with
> different types of clients, i.e I have a load of pentium 100s
> and some p2 500s which I would like to use together. Is this
> possible?
This is possible, but not really recommended. There's no real way to
ensure a certain part of a job gets run on a certain node, so you have
to assume all nodes are equal. But if you do this and have unequal
nodes, you'll find your jobs processing at the rate of the slowest
nodes.
I'd recommend just putting the 500's in the cluster, and if you really
need the extra power, put the 100's into a seperate cluster.
Jag
--ZARJHfwaSJQLOEUz
Content-Type: application/pgp-signature
Content-Disposition: inline
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.0.1 (GNU/Linux)
Comment: For info see http://www.gnupg.org
iD8DBQE5yj7F+pq97aGGtXARArFfAJ440pFNS3IRprQ4LCtjVKfJJmKMXACg3hJo
KU3RHY/5yHjtCpNjvf9lzeA=
=o3Se
-----END PGP SIGNATURE-----
--ZARJHfwaSJQLOEUz--
Fitch, Chester
As noted previously on this list, this all depends on your application. If
your application is tightly coupled, then yes, you will be limited by the
slowest compute node (unless you do some fairly sophisticated node profiling
and load balancing.)
One the other hand, if your application allows it, building a heterogeneous
cluster is a good way to put those old machines to productive work. MY
application, for example, is embarrassingly parallel - I'm running many,
many Monte Carlo simulations - each with differing input parameters. Each
simulation runs to completion on a single machine, and there is no processor
to processor communication required (other than between the Head and Compute
nodes). Therefore, I am able to utilize old hardware - what does it matter
(to the overall problem) if one node gets through 2 or 3 simulations in the
time it takes the slowest node to finish? My system is very small - only 7
compute nodes - consisting of 2 386 machines, 1 486 and 4 Pentiums (of
differing speeds), but I have been getting very good throughput,
considering...
Obviously, YMMV...
Just my $0.02
Chet
> -----Original Message-----
> From: Jag [mailto:agr...@linuxpower.org]
> Sent: Thursday, September 21, 2000 11:01 AM
> To: p.gri...@virgin.net
> Cc: beo...@beowulf.org
> Subject: Re: Beowulf and variable cpus
>
>
> On Thu, 21 Sep 2000, p.gri...@virgin.net wrote:
>
> >
> > Hi, I am new to Beowulf and have some questions,
> >
> > 1. Does anyone know if I am able to run a beowulf cluster with
> > different types of clients, i.e I have a load of pentium 100s
> > and some p2 500s which I would like to use together. Is this
> > possible?
>
> This is possible, but not really recommended. There's no real way to
> ensure a certain part of a job gets run on a certain node, so you have
> to assume all nodes are equal. But if you do this and have unequal
> nodes, you'll find your jobs processing at the rate of the slowest
> nodes.
>
> I'd recommend just putting the 500's in the cluster, and if you really
> need the extra power, put the 100's into a seperate cluster.
>
> Jag
>
>
_______________________________________________
Robert G. Brown
>
> Hi, I am new to Beowulf and have some questions,
>
> 1. Does anyone know if I am able to run a beowulf cluster with
> different types of clients, i.e I have a load of pentium 100s
> and some p2 500s which I would like to use together. Is this
> possible?
Sure, lots of ways. On embarrassingly parallel applications it is just
the total available CPU that counts, see e.g. the SETI or RC5DES
projects, the Stone Soupercomputer, and other heterogeneous efforts.
For other parallel applications that do involve a moderate degree of
synchronicity and communication, such a system can still be useful as
long as the slower machines don't act as a brake on the faster ones,
This generally means that you have to partition the work proportional to
the speed of the system and its capacity to communicate. For some kinds
of problems this is really not that difficult -- you just measure the
relative speed of the P5@100 to the P6@500 on your parallelized code
chunk (probably around 1:10) and give the PII's 10 times as much to do
between communication barriers (when everybody communicates). Fine tune
to correct for differential network speed and the larger communications
from the PII chunks (if any).
The programming IS more complex if the systems are heterogeneous,
though, as the balancing and tuning will likely have to be done "by
hand". You also have to have a really BIG load of P5@100's to make it
worth the hassle when one P6@500 (current cost of a 500 MHz Celeron node
is maybe $500-$600) can do the work of ten of the old Pentia. Just the
cost of the extra electricity (maybe 1-1.3 kilowatts times the time of
operation) can pretty much buy you the Celeron node over a year. It's
like paying to run a 900 Watt space heater in an air conditioned space
-- you pay twice. The space requirements, the extra switch/hub ports
required, and the extra human labor for maintenance and installation all
have to be considered as well. Finally, there is the inevitable drop
off in parallel efficiency as a job is distributed to more nodes.
Add this all up and you are probably better off economically NOT using
the Pentia and buying cheap but technologically current nodes of
equivalent total power by the time you factor in all of the real costs.
Still, there can be lots of circumstances that do make such a cluster
worthwhile -- low-budget/hobby beowulfs, beowulfs in schools with
strictly limited computer budgets (but somewhat elastic electricity/AC
budgets:-), clusters composed of your mix of Pentia and PII's sitting on
desktops and being ALSO used as workstations (somebody else pays for the
electricity and cooling and whatever is left of the CPUs is truly "free
computing power").
Hope this helps,
rgb
>
> Regards,
>
> Paul Grimshaw.
>
> -----
> Sent using MailStart.com ( http://MailStart.Com/welcome.html )
> The FREE way to access your mailbox via any web browser, anywhere!
>
>
> _______________________________________________
> Beowulf mailing list
> Beo...@beowulf.org
> http://www.beowulf.org/mailman/listinfo/beowulf
>
--
Robert G. Brown http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567 Fax: 919-660-2525 email:r...@phy.duke.edu