Slashdot Mirror
Ask Slashdot: "Be" is for Beowulf?
PhiberOptik asks: "Considering that
BeOS is so adept at
handling high I/O and network bandwidth, whould it make a
good OS for a Beowulf cluster? I know that since BeOS is
not open-source it would be harder to have applications
made, but they would be easier to maintain; as BeOS has a
standard set. Any ideas?" Could this be done? What OS
characteristics are necessary for a Beowulf cluster?
First of all, why would you want a Beowulf cluster of Be machines? The only thing I can think of that you would need any kind of clustering for BeOS is for distributed rendering, and most major studios just write their own software for handing off bits to Lightwave, etc..
Other than that, forget it, your dual-400MHz P2 is more CPU than you'll ever need, running BeOS. Hell, with personalStudio 1.0 from Adamation, you can do real-time video effects with _ONE_ 300MHz cpu.
I have a 400MHz P2 and I've never been able to use all the processor, other than mp3 encoding direct from cd.
--M
Beowulf clustered applications (along with COWs and many other distributed systems) are developed using mainly MPI or p-threads. There is almost no good reason to use PVM except to support legacy code.
And of the two, MPI is generally vastly superior to p-threads because (IIRC) MPI is a higher level implementation which provides communication routines that are optimized for each particular hardware implementation (depending on the version/implementation of MPI). For example,
the actual implementation of MPI_Reduce() will vary depending on whether the nodes are in a shared/non-shared memory environment -- in a non-shared environment (eg. a Beowulf cluster), MPI uses a tree-method in order to distribute the data among the nodes in parallel.
Anyways, the point is that MPI is really just a communications specification (Message Passing Interface) with language bindings (C and Fortran). What you really want is a set of client/server daemons that _implement_ MPI.
Okay, I'm done with the conch now.
Odds of being killed by lightning and winning the lottery in the same day: 1 in 2^55
The technology formerly known as "Wolfpack" (now NT Enterprise Cluster Server, or something like that) is intended for high availability (failover, etc). NT high performance clustering has been done, however. See, especially the "NT SuperCluster" at UIUC. There are a couple versions of MPI available for NT. I know of four, two no-longer-supported-but-still-available academic versions and two commercial versions derived from the academic versions. MPI-Pro (from MPI Software Technology, Inc.) and PateNT (from Genias) are the two commercial verions. Don't recall URLs, but MPI-Pro is advertised in Linux Journal, and if you're really curious, we all know how to use a search engine.
Christopher A. Bohn
cb
Oooh! What does this button do!?
Technically, you could not build a Beowulf from BeOS -- the definition that Don Becker, Tom Sterling, et.al., provide for "What is a Beowulf?" is that it must have a free OS (Linux, *BSD, other) so that, if need be, optimizations to the OS can be made. ... The limiting factor is getting interprocess communication libraries.
That said, there's no reason you couldn't set up a "Cluster of Workstations" (COW) which is basically what you're thinking of -- it's been done with Solaris, NT,
Christopher A. Bohn
cb
Oooh! What does this button do!?
The telnet server stuff has already been covered, but more importantly, people seem to have this untrue belief that the GUI in be is necessary. Most everything in the OS is launched via startup scripts, including the GUI. Granted, since no other interfaces are available, it would make life a bit difficult to use without Tracker/Deskbar running, it isn't impossible to comment out all the necessary lines to boot with no gui, and launch the necessary software in that script (telnet daemon, beowulf client, etc.)
seanf();
They should bother to read the entire post before commenting on it then.
Moderators also moderate down anything that is not pro-Linux/pro-GNU. "News for Nerds" does not mean "News for Linuxers". Anything that is pro-BSD is usually moderated down. Anything that is pro-Be is frequently moderated down.
It is not unusual to see random comments and junk marked "interesting" or "informative" regardless of the validity of the content. This is unacceptable behaviour the Slashdot moderators.
What does the fact that an operating system not being open source have to do with the ability to write applications for it? Just because something isn't open source doesn't mean it is any harder to develop for. I find developing under BeOS pretty easy, as a matter of fact. Nobody says you *have* to use the B class structure layed out for you. Prime examples would be the dozens of UNIX applications ported to the BeOS (ssh being a biggie).
I also don't see why you need to have a "single set" of software for an OS for use with Beowulf. I'd think you could handle installing the same software in your given cluster, and I don't know of software for Beowulf that people are wanting to distribute in a binary-only fashion. Do such beasts exist?
Don't get me wrong, I'm a Be fan, and I run it on my home box for weeks (sometimes while doing Linux technical support ;), but I don't see how it would be even as good as Linux for Beowulf purposes.
-josh
In practice, you can set up clusters on just about any system that supports both multithreading and networking. YMMV.
Um, the boxes that you are clustering will probably cost enough to make $50-$60 utterly insignificant, even if you use cheap celeries. My cost estimate for building a cluster of my own was about $400 US per node for just hardware (still haven't done it yet, in case you were wondering).
True. However, depending on the application, workarounds could suffice. A scheme that I speculated about would have a "lock memory region" command that you would call before accessing a data structure, that paged it in from the remote box that it was stored on if necessary, and also handled coherence. You would unlock it when you were finished with it. This would have substantial overhead, but for many classes of problem it would be adequate.
Similarly, for many classes of problem just a set of distributed clients with their own memory spaces would work. It all depends on what you are trying to do.
I've heard other people report problems with this chipset too. Send in a bug report to Be (check their web page). They'll either revise it themselves (if it was an internally written driver) or get the contractor to revise it if it was something that they paid for.
Re. the NE2000 compatible card, in principle that should work with the "NE2000" setting. In practice, who knows.
From a purely diagnostics point of view, I'd suggest trying each of the cards individually, to see if having the two cards in at once is causing problems (the uninitialized card may be causing bus or IRQ contention).
That assumes that you want to run the OS itself as a shared process. If all you want is a distributed application, this isn't necessary (though kernel support of shared memory helps, as another poster pointed out).
I was here before the present moderation system was in place. It wasn't pretty. While some posts are unjustly moderated down (or up), the vast majority of moderation that I've seen has been deserved.
In this case - I agree that it was an informative post. However, the moderator that marked it down didn't agree. If other moderators think that it was informative, it will be marked up again.
Last but not least, the moderators are the _readers_. Read the moderator guideline documents, and the articles on the subject. If you browse slashdot regularly, and post, you _will_ get moderator access now and then. The whole "all moderators are corrupt" view looks a bit strained in this light.
There are some special purpose applications where an OS can make a real difference. Traditional operating systems, for example, have trouble keeping up when there are a lot of tiny chunks of data that need to be processed by user-level server processes (one of the reasons why routing, logging, and nfsd are in the kernel). There are some operating systems specifically designed for such applications, but they have to make other trade-offs. And I don't see this being an issue with Beowulf applications.
Marketeers like to make claims about one system being a lot faster than another (c.f. Windows NT vs. Linux), but in practice, I find those are pretty meaningless.
For Beowulf in particular, given that it's being developed with Linux and given that the people on the project know very well how to hack the Linux kernel and networking code, even if there were a bottleneck in some release of Linux, it's a good bet that it would get fixed really quickly. How much effort do you think Be Inc. is going to invest in addressing performance problems for distributed scientific computing, given that their target market is desktop multimedia?
There is SockHop for BeOS.
Programs have to be written specificaly for SockHop and use its base class instead of BLooper (Be's basic thread + event loop class) SockHop also uses a set of classes to handle routing of BMessages between threads.
I know that since BeOS is not open-source it would be harder to have applications made, but they would be easier to maintain; as BeOS has a standard set.
A standard set of what? If you are talking about system calls, I wouldn't call Linux "non-standard". In any case, "proprietary" is the opposite of "standard".
---
Put Hemos through English 101!
Linux MAPI Server!
http://www.openone.com/software/MailOne/
(Exchange Migration HOWTO coming soon)
"I know that since BeOS is not open-source it would be harder to have applications made"
What makes you think that open-source software is any easier to develop? Be has documented the entire BeOS API and provides tons of sample code on every CD they sell. If you subscribe to the Be developer newsletter then they usually send more sample code plus tutorials in every issue. Don't get me wrong, I am not anti-linux. I do question the logic behind GPL'ing apps all the time. As a person trying to learn C++, I want to be able to make money off of the app itself... not supporting it. OSS has its place, and it is definitely got an impressive list of apps. However we need moderation because an all OSS world is no different than an all closed source world.
---Got Coffee?---