Mailing Lists: Apple Mailing Lists
Image of Mac OS face in stamp
 
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Re(2): Difference Between Xgrid Cluser and a Beowulf

> A lot of this depends on what you call a Beowulf.  If you ask a dozen
> people, it seems you'll get about 14 different answers.

Very true, and each of us will be supremely overconfident that ours is the only real answer. Here's mine, based on a few years of helping to build clusters, cycle stealing networks, and grids. A longer set of notes I put together a few months ago is at: http://www.bioteam.net/cdwan/grid.pdf

The Beowulf project was intended to create a low cost alternative to large supercomputers. One of its central goals (long since set aside by most people) was to provide a single system image. I consider that to be the hallmark of a "beowulf" style cluster: It is a tightly integrated system with fixed topology which the user interacts with as a single large machine. Traditionally, the interconnect on a machine of this style is as fast as possible, since it is the closest we'll come in this conversation to a traditional parallel processing supercomputer.

A compute farm relaxes the single system image requirement. Instead of the user experiencing a single multiprocessor system (with, perhaps, a scheduler buried in that single OS), each node runs its very own installation of the OS. The user submits jobs to a distributed resource manager (DRM) or queuing system, which schedules the jobs onto sets of compute nodes using daemons (and perhaps a shared filesystem or database) for interprocess communication. Compute farms are ideal for trivially parallel operations including monte carlo simulations, animation rendering, and sequence based bioinformatics. DRMs originally designed for this style of machine include PBS, PBS Pro (Now "Torque"), Sun's Grid Engine (SGE), Platform's LSF, and others. Xgrid can be used to manage a rack full of Xserves, but there are better options if those machines are to be dedicated cluster nodes. There are free and commercial distributions of cluster software for this style of resource.

Networks of workstations and cycle stealing environments are distinct from compute farms, though they are appropriate for the same class of problem. Instead of the compute nodes being dedicated to the cluster, they do double duty at other tasks, perhaps being workstations or part-time servers. Internode communication and scheduling become truly difficult as we accept that building and campus networks are part of the data motion puzzle, and user's lunchtime is a part of scheduling. Condor, from the University of Wisconsin at Madison is the most mature system for constructing cycle stealing environments. Seti@home (and its imitators) popularized the style. United Devices and Platform each sell software to implement this sort of resource. And, of course, Xgrid fits this definition.

Grid computing, as defined by Ian Foster and Carl Kesselman, and as discussed at the Global Grid Forum (Sun and Apple's co-opting of the definition aside) is distinguished by the fact of resource sharing (compute, network, authorization, and data) across organizational boundaries. Theirs is a three part definition:

* A grid connects non-uniform, physically separated resources which are not under any single administration or ownership
* A grid uses open protocols
* A grid accomplishes nontrivial work.

In my opinion, if a single enterprise funds and builds the resource, and there is a single software platform enabling it, and there is a single administrative password which can be used to make central decisions about global changes, we're talking about a wide area cluster or something like that. Grids are what happens when clusters begin to interoperate across administrative boundaries (to enable virtual distributed organizations and leverage blah blah blah...). SGE, LSF, Torque, Xgrid, Globus, Web Services, and Napster have all been used to create grids which fit the definition.

In my opinion, a grid is distinguished primarily by social constraints, not technical ones. Not one of the above technologies will, by definition, give you a grid after you install it.

Also SUN is calling their cluster software "Grid Engine". Must there be a
difference at all? Maybe "Grid" is just a more common name nowadays than
"Beowulf" (for marketing reasons) and in the end it's describing the same
technics (from the point of hardware).

For the sake of clarity, I support using some sort of unique term for each of these very different styles of computing. Otherwise we get to keep having this conversation over and over.

-Chris Dwan
The BioTeam (sent from my old University of Minnesota address because I haven't updated the list subscription with my new email yet. Sorry for any confusion this may cause)
_______________________________________________
xgrid-users mailing list | email@hidden
Help/Unsubscribe/Archives: http://www.lists.apple.com/mailman/listinfo/xgrid-users
Do not post admin requests to the list. They will be ignored.

References: 
 >Re: Difference Between Xgrid Cluser and a Beowulf (From: "Jay A. Kreibich" <email@hidden>)
 >Re(2): Difference Between Xgrid Cluser and a Beowulf (From: email@hidden (Reuti))



Visit the Apple Store online or at retail locations.
1-800-MY-APPLE

Contact Apple | Terms of Use | Privacy Policy

Copyright © 2007 Apple Inc. All rights reserved.