sam@lfcs.edinburgh.ac.uk (S. Manoharan) (05/26/89)
I am interested in getting some statistics on real parallel programs. Particularly, given a dependency graph of a parallel program, I want to know: 1) the times spent in executing the nodes, and 2) the amounts of information (say, in bytes) transmitted to the successor nodes upon completion. A compiler can be made to generate such information, at least for the programs that do not depend on runtime data. I would like to know whether someone could help me getting data of this kind, or a compiler/parser to generate such data. All pointers are welcome. Thanks.
eugene@eos.arc.nasa.gov (Eugene Miya) (05/31/89)
In article <5605@hubcap.clemson.edu> sam@lfcs.edinburgh.ac.uk (S. Manoharan) writes: >I am interested in getting some statistics on real parallel >programs. Wow! You aren't asking too much ;). Lots of open research questions here (what constitutes a stat? do you want the data flow graph? the control flow graph, dependences, etc.). A Lot of stuff and we really don't know how to interpret it either. You should post your findings (be they negative or positive) for the general net. Good luck! --eugene
cline@sun.soe.clarkson.edu (Marshall Cline) (06/09/89)
In article <5605@hubcap.clemson.edu> sam@lfcs.edinburgh.ac.uk (S. Manoharan) writes: >I am interested in getting some statistics on real parallel >programs. Particularly, given a dependency graph of a >parallel program, I want to know: > 1) the times spent in executing the nodes, and > 2) the amounts of information (say, in bytes) transmitted to > the successor nodes upon completion. >A compiler can be made to generate such information, at least >for the programs that do not depend on runtime data. Check out last month's (May 1989) IEEE Transactions on Computers: Helmar Burkhart & Roland Millen, "Performance-Measurement Tools in a Multiprocessor Environment", IEEE Trans Computers, 38,5 (May 1989), 725-737. The authors describe a multiprocessor (as opposed to multicomputer) environment for which they have built/are building performance measuring tools. The introduction is quite long, such that the paper doesn't assume a vast amount of previous knowledge on the subject. Their tools measure the performance loss due to a number of factors, including contention over the system bus (remember: theirs is a multi_PROCESSOR_ system), idle processor loss, access conflict over shared resources, access avoidance even in the absence of conflict requires additional overhead, "braking" loss (time spent waiting for other processors to halt when I've already finished the problem), etc, etc. Hope this helps! Marshall -- ________________________________________________________________ Marshall P. Cline ARPA: cline@sun.soe.clarkson.edu ECE Department UseNet: uunet!sun.soe.clarkson.edu!cline Clarkson University BitNet: BH0W@CLUTX Potsdam, NY 13676 AT&T: (315) 268-6591