jmm@uoregon.UUCP (11/30/87)
I would appreciate any help finding sources on Connectionism and the Connection Machine. Thank you in advance. malcolm jmm@uoregon.edu
pase@ogcvax.UUCP (Douglas M. Pase) (11/30/87)
A good place to start is: D. Hillis and G. L. Steele, "Data Parallel Algorithms," Communications of the ACM, Vol. 29, No. 12, December 1986, pp 1170-1183
grunwald@uiucdcsm.cs.uiuc.edu (12/01/87)
The connection machine is not connectionism. A reference for the connection machine is "The Connection Machine" by Hillis, MIT Press. A reference for connectionism is "Parallel Distributed Processing" by Rummelhart (I think the spelling is correct). It would be better to ask for references concerning this in comp.ai.
reiter@endor.harvard.edu (Ehud Reiter) (12/11/87)
The Connection Machine is usually presented as 64,000 1-bit SIMD processors hooked up with a hypercube router. However, from talking to a few people who have experience with it (the CM2, not the CM1), I get the impression that many applications basically ignore the 1-bit processors and the hypercube router (which is very slow - it takes 1 ms to send a message), and instead do all their computing with the 2000 Weitek FPU's, and do all their communicating with the "NEWS" system, which basically is a fast 2-D mesh interconnect. That is, the alternative model of the Connection Machine which these applications use is that it is a 2-D mesh of 2000 fast FPU's, backed by 512MB of memory and a very high bandwidth disk I/O system. Like a modern-day version of the Illiac IV, I guess (but one that actually works, unlike the Illiac). To me, anyways, this "Illiac" model seems much more useful (for getting real work done) than the "hypercube" model. My question is, how do people who are running real applications on the CM2 use it? Does anyone who actually has a CM care to comment? I realize that there is one serious application, the news-wire retrieval program, which does use the 1-bit CPU's to do its communicating (although it doesn't use the hypercube router very much) - are there other (commercial) applications which use either the 1-bit CPU's or the hypercube router? Ehud Reiter reiter@harvard (ARPA,BITNET,UUCP) reiter@harvard.harvard.EDU (new ARPA)
ortiz@think.COM (Luis F. Ortiz) (12/11/87)
In article <3516@husc6.harvard.edu> reiter@harvard.UUCP (Ehud Reiter) writes: >The Connection Machine is usually presented as 64,000 1-bit SIMD processors >hooked up with a hypercube router. However, from talking to a few people who >have experience with it (the CM2, not the CM1), I get the impression that many >applications basically ignore the 1-bit processors and the hypercube router >(which is very slow - it takes 1 ms to send a message), and instead do all >their computing with the 2000 Weitek FPU's, and do all their communicating >with the "NEWS" system, which basically is a fast 2-D mesh interconnect. First, I would like to correct something. The CM-2 has no 2-D mesh interconnect between processors like the CM-1 does. It turns out that the hypercube connections between processors are suffcient to embed a N-D lattice in the M dimensional hypercube (where N < M). >That is, the alternative model of the Connection Machine which these >applications use is that it is a 2-D mesh of 2000 fast FPU's, backed by >512MB of memory and a very high bandwidth disk I/O system. Like a >modern-day version of the Illiac IV, I guess (but one that actually works, >unlike the Illiac). To me, anyways, this "Illiac" model seems much more >useful (for getting real work done) than the "hypercube" model. >My question is, how do people who are running real applications on the CM2 >use it? Does anyone who actually has a CM care to comment? I realize that >there is one serious application, the news-wire retrieval program, which does >use the 1-bit CPU's to do its communicating (although it doesn't use the >hypercube router very much) - are there other (commercial) applications >which use either the 1-bit CPU's or the hypercube router? > Just to mention a few of the projects that we have worked on: 1) FFT's: It turns out that you can directly use the hypercube communications structure of the CM to great advantage while still looking at the machine as a one dimensional sequence of processors. 2) 3-D Air flow: Here we map a 3-d lattice onto the hypercube and use each processor to simulate a vlume of air. 3) QED: Quantum electrodynamics can map nicely into a 4 torus with each processor keeping track of the field strength in eight directions. 4) Cellular Automata: This kind of a research application uses almost no (if any) floating point, and spends its time doing nothing but 1-bit operations interspersed with aribtrary dimensional lattice communications. I guess my real point is that the machine has a special features (like fast 1-bit arithmetic, fast floating point, and hypercube connectivity, that are used differently by different applications. The field of parallel processing is still young and, quite frankly, there is still a lot left to learn. When people try to use the machine, oftentimes they one idea as to how the problem is to be solved, but end up implementing it in a totally differnt manner. Often you are pleasantly surprised by something you though was difficult, turning out to be elegantly simple (like diffusion equation solutions). I think that viewing the CM-2 as being a 2-D mesh of floatinbg point processors is too limited. ---------------------------------------------------------------- Luis F. Ortiz Thinking Machines Corporation ARPANET: ortiz@think.com, ortiz@yale.edu UUCP: {decvax,harvard,seismo,cmcl2}!yale!ortiz, harvard!think!ortiz BITNET: ortiz@yalecs.BITNET "Whenever people agree with me, I always think I must be wrong." --- Oscar Wilde ----------------------------------------------------------------
Am I consing yet?) (12/15/87)
>Posted by: decwrl!hplabs!hp-sdd!ucsdhub!sdcsvax!ames!ll-xn!husc6!endor!reiter > >The Connection Machine is usually presented as 64,000 1-bit SIMD processors >hooked up with a hypercube router. Who is to say for sure? How is this different from a single processor that happens to have a 64k-bit word size, some slick masking commands, and some hellacious shift intructions? -Jeff Bell
cs3ia3ac@maccs.McMaster.CA (Howard Betel) (01/21/89)
I'm not a regular reader of this group (here comes trouble) so if my
question belongs somewhere else then flame away, but do it through
mail so we don't tie up the net.
I'm currently a third year Comp Sci student at McMaster University
in Hamilton (Canada) and I am taking a computer architecture course
this semester. One of the major projects of the term is a 25 page
report an a half hour seminar on an architecture related field (that
sure is wide open). I've chosen the Connection machine but am finding
it hard to get info. Can anyone point me to some current articles or
reports that I might be able to get my hands on? ANY help would
be appreciated.
-------------
Howard Betel cs3ia3ac@maccs.UUCP
cs3ia3ac@maccs.McMaster.CAspector@vx2.NYU.EDU (David HM Spector) (01/28/89)
You might take a look at:
The Connection Machine
D. Hillis
MIT Press (I Think, my copy is at home)
_DHMS
-------------------------------------------------------------------------------
David HM Spector New York University
Unix Manager/Senior Systems Programmer Stern School of Business
ARPAnet: SPECTOR@GBA.NYU.EDU Academic Computing Center
USEnet:...!{uunet,rocky,harvard}!cmcl2!spector 90 Trinity Place, Rm C-4
HamRadio: N2BCA MCIMail: DSpector New York, New York 10006
AppleLink: DHMSpector CompuServe: 71260,1410 (212) 285-6080
"Capital punishment is our society's recognition of the sanctity of human life"
- Senator Orrin Hatchjps@cat.cmu.edu (James Salsman) (08/09/89)
In article <2238@jhunix.HCF.JHU.EDU> ins_atge@jhunix.UUCP (Thomas G Edwards) writes: > And just because every processor gets the same instruction feed, one must not > think that every processor is "doing the same thing." Each CM processor > can hold an index to an array located in that processor, Only on the CM-2. Also, don't forget the condition sense flag lines of the nanoinstruction stream. > so with the right software, the CM could become a MIMD machine. Please do not attempt this. I have tried for months to optimize a MIMD-emulator on the CM-2 with very disappointing results. It might be a fun class project for a bunch of novice hackers learning parallelism. :James Disclaimer: They tell me that I'm not allowed to post to the net anymore! This is totally bogus but if they decide to remove my account, there's nothing I can do about it. Fascists. -- :James P. Salsman (jps@CAT.CMU.EDU)