thomasf@ius3.ius.cs.cmu.edu (Thomas Fahringer) (06/24/91)
I am parallelizing a couple of applications on the Intel IPSC/860. Now I want to do some performance analysis and prediction. For this reason I am desperately looking for some trustable performance formulas and information about the Intel IPSC/860 for following purposes: 1. the time it takes to transfer a message of n (varying) bytes from a source to a destination node (single and multiple hops). 2. slow down factor of a transfer operation (according to point 1.) due to channel contention. 3. communication primitives behavior for different message sizes and different number of hops to go between source and destination node. I am mainly interested in following communication primitives: - blocking sequential (csend/creceive) - blocking full duplex (csend/creceive) - non-blocking receive, blocking send - blocking receive, non-blocking send - fully non-blocking Which communication primitive should I prefer ? 4. any formula describing the slow down behavior of an array access within a FORTRAN DO-loop induced by cache misses. I am interested in any kind of cache influnce in my computation on the IPSC/860. I am well aware of cache technology, but have no information about the Intel IPSC/860 cache architecture, strategies, etc. 5. Is there anyone out there who has measured performance of all the standard and intrinsic operations on the Intel IPSC/860 ? 6. Is there another bboard containing more information about the Intel IPSC/860 than this one. 7. If someone already collected a couple of IPSC 860 summaries, please forward them to me. I post a summary if anyone requests. Eternal thankful, Tom
berryman-harry@CS.YALE.EDU (Harry Berryman) (06/24/91)
A good summary of communications costs can be found in the study by Shahid Bokhari. It is "Communications Overhead on the Intel iPSC/860 Hypercube" ICASE Interim Report 10. Bokhari did extensive studies of the multihop communication costs under a variety of situations including contention for nodes (which made little difference) and contention for edges (which made a lot of difference). Scott Berryman Yale University Computer Science Department and ICASE/NASA Langley Research Center
jet@karazm.math.uh.edu (J Eric Townsend) (06/25/91)
I can answer at least one of these... :-) In article <1991Jun24.122022.19173@hubcap.clemson.edu> thomasf@ius3.ius.cs.cmu.edu (Thomas Fahringer) writes: >1. the time it takes to transfer a message of n (varying) bytes from a > source to a destination node (single and multiple hops). >3. communication primitives behavior for different message sizes and > different number of hops to go between source and destination node. If message_size > 100 bytes, a "is there room to send this message" message passing routine happens. You can avoid this by using the FORCE message type and writing code that insures the receiver always has room. >6. Is there another bboard containing more information about the > Intel IPSC/860 than this one. This is probably the best place. Also, I'm building an ftp site for ipsc related stuff, so feel free to drop data into karazm.math.uh.edu:pub/Incoming. (fyi: This is a newsgroup, not a "bboard".) -- J. Eric Townsend - jet@uh.edu - bitnet: jet@UHOU - vox: (713) 749-2126 Skate UNIX! (curb fault: skater dumped) PowerGlove mailing list: glove-list-request@karazm.math.uh.edu