Thanks to everyone for their response on the Big SMP thread. I apologize I wasn't able to respond earlier, but school--final's week in particular--has been disctracting me. ;-) I'll now answer some of the questions people have asked me, as well as to summarize what I've read and been told since my original post. I don't have any problem with getting up to speed on PVM or some MPI variant--but I'm just an energetic headstrong youth, and besides it's my advisor's grant money (well, will be if the proposal goes through). He (Andrew) wants shared memory--whether UMA or NUMA doesn't matter so much. So I'll report to him the number of people suggesting clusters, but I think he's selectively deaf on this issue. However, someone mentioned a package called "dipc" which would transparantly (?) share memory on a non-shared architecture, i.e. clusters. Can anyone point me to details about dipc or similar programs, and share any experience? I've read the Parallel Processing HOWTO by the Purdue PAPERS guy (sorry I don't recall his name) which mentioned several of these, but it was lacking details. Experience counts for a lot when your advising someone on how to spend their grant money. So with Linux, shared-nothing (or very little) clusters are really the only option for lots of processors and memory. I'm going to summarize below the non-linux options people have suggested to me on this group, and I hope it's not (too) inappropriate for this particular newsgroup. Besides, it seems that many people here like talking about big computers when someone else's money is involved! ;-) Also, my email address is [EMAIL PROTECTED] for any formal complaints about this divergence from Linux, or any help on the machines below. SGI ORIGIN: Big, expensive, deluxe. Third party memory available, and using third party disks as well could greatly defray the cost. I can't believe how much they charge for memory--I think the quotes are confidential so I won't mention the exact price. But it seems that when things get pretty serious, SGI salespeople can make some big price reductions. They don't like hearing about 3rd party products, of course, even if that's what it will take for us to buy their machine. The salesman told me about an Origin system(s) catching on fire because of third-party memory--makes a good story in the very least. Other details about Cellular IRIX include that the POSIX threads implementation doesn't allow threads to migrate from one processor to another. IBM S70,SP2: Not much mentioned about IBM, and I'm wondering why. Of particular interest to us are the processors, which are the only non-Intel processors I'm aware of with faster integer performance than floating-point. For our application, described at the bottom, integer performance is more important. Information about how well AIX handles the shared-memory bit and threads for a multiprocessor architecture would be appreciated. Again, if nobody else on the newsgroup wants to know about this, my email address is [EMAIL PROTECTED] Especially desired is real info on this mysterious rsII-64 cpu they put in the S70. SUN ENTERPRISE: I haven't heard a whole lot about the Enterprise servers either, and the SUN rep I'm supposed to talk to has been difficult to reach. Any info/experience on big SUNs? Again, [EMAIL PROTECTED] if needed. SEQUENT: I haven't really looked into this. It seems unnatural to use Intel processors for the kind of machines they build, but what do I know? INTEL: Of course clusters (and Sequent) are the only options here. It seems everyone is screaming "More memory, please!". I started reading about the memory architecture in protected mode, and scanned some of the memory handling code for Linux (setting up for the switch to protected mode). It's like a comic book, laughs on every page. What a wacky processor family. x86 should be relegated to Horowitz and Hill's Bad Design Ideas, in _The Art of Electronics_. ALPHA: Cool cpu, inspired me in my computer architecture course. Still inspires me. But the motherboards really seem lacking. The best advice I've been given so far was from Alan Cox: get free lunch out of the deal from all the fawning salespeople. Unfortunately, IBM's salespeople are not in town, but SGI and SUN's are! And finally, what we're going to be doing with such a machine: We don't exactly know yet. This is a research setting, and we're outgrowing everything we can get our hands on so we're looking ahead. The application is large-scale data mining. For instance, the Sloan Sky Survey will be producing about 30 or 40 terrabytes of raw (ascii) data. Astronomers collect the data, do some minor processing, and store it. Then some inconsiderate astronomer wants to use the data for something, maybe to check his "gosh, I misplaced all the black matter" model against real data. So the astronomer wants to run queries on the data, maybe thousands of queries. For the dataset sizes we're looking at now (in the pathetically small size for the overall Sloan survey, but I don't have real numbers handy right now and I don't want to make them up), our data mining software is barely holding on using single processor Intels with maybe 1/2 gig of memory. The good news is, according to Andrew, all the competing software (this includes IBM, etc) died long ago on these datasets. So we've got to improve our algorithms, but also make better use of bigger hardware. We're single-threaded, single-process right now, but that should change when we get the 4 cpu Xeon in a month or so. The big machine we've been talking about above is farther down the road, but hopefully no more than 6 months--I try to stay out of the administrative details right now, since I don't want to gain the responsibility of actually writing the grant rather than just spending the money. Anyway, we need to have a machine and sound reasoning for the grant proposal. None of us have any hardcore parallel experience, though this may be the mainstay of my work next sememster. Andrew's research group goes beyond all this into other processing issues. If you're interested, the homepage is www.cs.cmu.edu/~AUTON. Thanks again for all of the help! -Paul Komarek [EMAIL PROTECTED] - Linux SMP list: FIRST see FAQ at http://www.irisa.fr/prive/mentre/smp-faq/ To Unsubscribe: send "unsubscribe linux-smp" to [EMAIL PROTECTED]
