home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.parallel
- Path: sparky!uunet!wupost!darwin.sura.net!gatech!hubcap!fpst
- From: gusciora+@cs.cmu.edu (George Gusciora)
- Subject: Pipeline performance analysis
- Message-ID: <BuG0nu.Epq.1@cs.cmu.edu>
- Sender: news@cs.cmu.edu (Usenet News System)
- Nntp-Posting-Host: n3.sp.cs.cmu.edu
- Organization: School of Computer Science, Carnegie Mellon
- Date: Sat, 12 Sep 1992 02:09:28 GMT
- Approved: parallel@hubcap.clemson.edu
- Lines: 36
-
-
- I have a system which can be modeled by a large pipeline where the
- stages of the pipeline perform an amount of computation which is dependent
- on the particular data set which is passing through that stage (stages do
- not perform the same amount of computation on each data set). I'm
- looking for some performance analysis work on similar problems. If
- you are familiar with this, please read on.
-
- The basic overview of the model is this: each stage of the pipeline
- performs the same amount of work over all of the sets of data
- which must be processed. However, the amount of computation done
- at a stage varies according to which data item is flowing through
- that stage.
-
- A bit more precise (in case anyone wants to discuss this)
- we consider a P-stage pipeline which is used to process n items of
- data. Label the amount of computation to be done on data item i at
- stage j as Eij. Between stages in the pipeline are large data buffers
- which, for simplicity, we can assume are arbitrarily large. The
- pipeline is 'balanced' in such a way that for any stage j, the
- sum E0j + E1j + ... Enj is equal to a similar sum constructed
- at any other stage.
-
- I would like some performance analysis of a system like this. My guess
- is that it will depend on the range and distribution of E*j for each stage.
- Does anyone know of work done in analyzing the performance of systems
- similar to this? Since this is such a specific topic, please send replies
- to me (gusciora@cs.cmu.edu) and I'll post a summary if people are
- interested.
-
- Thanks,
- George Gusciora
- gusciora@cs.cmu.edu
- Department of Electrical and Computer Engineering
- Carnegie Mellon University
-
-