[Beowulf] SGI to offer Windows on clusters

Robert G. Brown rgb at phy.duke.edu
Fri Apr 13 06:15:11 PDT 2007


On Fri, 13 Apr 2007, laytonjb at charter.net wrote:

> rgb,
> Perhaps I'm not thinking as broadly as Rich. But I see a web-base solution as a
> better idea than asking (or forcing) ISV's to put some new code in their applications
> to run on a cluster (BTW - in my experience some customers use the GUIs that
> come with ISV codes and some don't.).

I agree, in detail -- I just think that what you are describing is
something that has been implemented many times over in grid-land,
specifically (to my knowledge) in ATLAS (the DOE/HEP project, not the
linear algebra library).  See e.g.

   http://atlas-proj-computing-tdr.web.cern.ch/atlas-proj-computing-tdr/Html/Computing-TDR.htm

But elsewhere as well in e.g. bioinformatics land and chemistryville.

Note that the "grid" in question is basically a union of clusters
capable of running anything from EP codes (which is probably the rule
for obvious reasons) to PVM/MPI real parallel codes.  The clusters are
"generic", that is and could be designed or built to provide any level
of cluster service or be optimized for any particular code type.

The thing that makes these clusters into a grid is that one agency pays
for them all and insists that the resource not be wasted, so that any
researchers associated with the (globe-spanning) project can access idle
resources in any cluster.  Since these researchers have the usual wide
range of desktop workstations, a web based interface is essential, but
there are strong requirements on the interface beyond that -- strong
authentication, encryption etc.  Since the web portals are of necessity
open to the entire internet they have to be secure enough that local
cluster admins don't feel like they are opening their internal network
to the wolves by putting the interface up at all.

Internally, the clusters are architected much as you describe, with a
shared disk resources that for the proposal I helped to work on needed
to START at the ~100+ TB level -- HEP likes BIG datasets -- and
commensurately very high bandwidth on the primary ATLAS site PoP -- we
were talking about colocating on the lambda rail and the internet2
backbone, for example.  With scaleability to eventual petabyte capacity.
The internel IPC network was actually more modest as the code base
expected to be run was mostly EP.  The disk service network was where
one would likely invest in higher speed and better performance networks
or better data delivery designs.

The only point I was really making is that this is a good problem not to
reinvent wheels for.  Microsoft is even a Globus sponsor, so one hopes
that they are not openly averse to a Globus interface to linux-based
clusters embedded in a MS-based workstation environment.  Or rather,
that while their marketing department might be appalled at the very
suggestion (when one could equally easily pay them large amounts of
money for their native clustering product and support thereof) it will
work just fine.  A grid toolkit isn't just for "grids".  Indeed, one
could spend many pages just debating where a "cluster" stops and starts
becoming a "grid".  I personally think of a grid as a union of clusters
with a portable interface enabling authenticated, policy-enabled,
resource-matched WAN access to any resources within the union, but
obviously a set of one is still a set -- the WAN interface ultimately
defines a "grid".

Arguments, corrections, further discussions welcome as always.  And I am
NOT asserting that ATLAS's model should be naively emulated, BTW -- the
design sucked in several ways IMO when I went through it before, in
particular its reliance on pacman and the usual (at the time) wierdness
in the HEP community in insisting on the use of archaic versions of
linux in order to guarantee that their antique and difficult to rebuild
codes would find tested libraries in place.  There was a clear conflict
between being able to support modern devices with a modern kernel and
its drivers in a modern distro and this "we haven't ported or tested
past RH 7.3" mentality that may have finally disappeared with e.g.
Scientific Linux representing a much more up to date and portable and
rebuildable version of much of e.g. the CERN code base and other tools,
but I suspect that part of it is still there.

ATLAS members on the list?  Comments?

    rgb

>
> Jeff
>
>
>> On Thu, 12 Apr 2007, laytonjb at charter.net wrote:
>>
>>> I really think the web interface is the way to go. This way you can submit jobs from
>>> any machine that has a browser (Linux, Windows, Mac, etc.).
>>
>> Isn't that what gridware basically does already?  Doesn't SGE provide a
>> web interface to resources (a cluster) it controls?  Isn't that a
>> significant part of the point of the Globus project?  The ATLAS grid
>> (IIRC) uses a grid interface, more or less, to provide just this sort of
>> layer isolation between cluster/grid resource and the user.
>>
>> There are problems with this, of course.  It is wonderful if the
>> grid/cluster already has a canned package installed, so that what the
>> user "submits" is a parametric dataset that tells the cluster how and
>> what to run with that package.  BLAST etc work fine this way and there
>> exist cluster/grids architected just this way for this purpose, from
>> what I can tell via google.  If you want to run YOUR source code on the
>> cluster from a different OS, however, well, that's a problem isn't it?
>>
>> I think that there have been efforts to resolve even this -- ways of
>> submitting your source (or a functional binary) with build (or run)
>> instructions in a "package" -- I vaguely remember that ATLAS attempts to
>> implement a little horror called "pacman" for this purpose.  I leave to
>> your imaginations the awesome mess of dealing with library requirements,
>> build incompatibilities, mistaken assumptions, and worse across
>> architectures especially ones likely for people who write in MS C++ (or
>> a C downshift thereof) and expect it the source to "just run" when
>> recompiled on a linux box.
>>
>> Practically speaking, for source code based applications if the user has
>> a linux box (or even a canned vmware linux development environment they
>> can run as a windows appliance -- and there are many of them prebuilt
>> and available for free so this is no longer that crazy a solution on a
>> moderately powerful windows workstation -- and sufficient linux
>> expertise to work through builds thereupon, they can develop binaries or
>> build packages that they can submit to a cluster via a web interface
>> that hides all cluster detail.  If not, then not.
>>
>> Joe of course is building specific purpose clusters for many of his
>> clients and hence can successfully implement either canned software
>> solutions OR can manage the porting, building, preinstallation of the
>> client's software so that they can use it via a web-appliance interface.
>> Basically they purchase his expertise to do the code migration -- which
>> is again fine if the source is mature and unlikely to need a lot of
>> real-time tweaking and if they mostly want an appliance with which to
>> process a very large data space or parametric space a little at a time
>> (so "jobs" are parametric descriptions used to start up a task).
>>
>> There are various other details associated with gridware and cluster
>> usage of this sort that make the idea "good" or "bad" per application.
>> If the application is bottlenecked by data access -- it processes huge
>> files, basically -- one can spend a lot of time loading data onto the
>> cluster via e.g. the web interface compared to a little time running the
>> application on the data, something that can perhaps be done more
>> smoothly and faster with a native shared disk implementation instead of
>> double hits on native disk on both ends plus a (probably slow) network
>> transfer.  Accessing other resources -- GUI access to the program being
>> run, for example -- similarly depends strongly on having the right hooks
>> on both ends.
>>
>>     rgb
>>
>>>
>>> Jeff
>>>
>>>> Here is a proactive suggestion for keeping open source
>>>> ahead of Microsoft CCS:
>>>> 1. I think CCS will appeal to small shops with no prior cluster
>>>>     and no admin capability beyond a part time windows person.
>>>> 2. such customers are the volume seats for a range of desktop
>>>>     CAD/CAE tools.
>>>> 3. Such ISVs will see potential of license growth, and will
>>>>     likely choose to tie-in to the Microsoft message of ease-of-use.
>>>>     A big feature here, in my view, is the one-button-job-launch.
>>>>
>>>> This means, for Linux to have a position as the backend
>>>> compute cluster, we must have this one button job launch
>>>> capability.  A Windows library must be available to
>>>> the ISV, to provide a job submission API  to the batch
>>>> scheduler.  With such a feature, the ISVs can be
>>>> persued to incoporate.
>>>>
>>>> Ideally the job submission API is a kind of standard, so
>>>> the ISV does not see duplicate work versus the batch scheduler
>>>> used.
>>>>
>>>> So,
>>>> a) we need a job submission API, and
>>>> b) we need the Windows library added to Linux batch schedulers.
>>>>     (I'm not saying the scheduler runs on Windows, we just need
>>>>     the submission/retrieve portion).
>>>>
>>>> Does such exist already?
>>>>
>>>> Thanks, Rich
>>>> Rich Altmaier, SGI
>>>> _______________________________________________
>>>> Beowulf mailing list, Beowulf at beowulf.org
>>>> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
>>>
>>> _______________________________________________
>>> Beowulf mailing list, Beowulf at beowulf.org
>>> To change your subscription (digest mode or unsubscribe) visit 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:rgb at phy.duke.edu
>>
>>
>

-- 
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:rgb at phy.duke.edu





More information about the Beowulf mailing list