Today's Events (reminder)


Date: Mon, 30 Oct 2000 15:35:12 -0600 (CST)
From: David Melski <melski@xxxxxxxxxxx>
Subject: Today's Events (reminder)
Hey all, 

This is just a reminder that I'm giving the PL seminar today, at 4:00 in
room 1325.  

Hope to see you there,
Dave


---------- Forwarded message ----------
Date: Mon, 30 Oct 2000 06:00:18 -0600 (CST)
From: CS Dept. Talks <colloq@xxxxxxxxxxx>
To: msgs-cs@xxxxxxxxxxx
Subject: Today's Events


4:00 pm, 1325 CS
PL Seminar:  David Melski, UW, "Profile Driven Program Analysis"

        Ammons and Larus presented a technique for  improving  the  results  of
   daqa-flow analysis along hot paths [Ammons-Larus, PLDI'96].  In this talk, I
   will present an extension of their work based on Melski and Reps'  interpro-
   cedural  path-profiling  technique.   Rather  than using the intraprocedural
   path-profiling technique of Ball and Larus, we use a technique that collects
   information  about  interprocedural  paths  (i.e., paths that may cross pro-
   cedure boundaries).  Hot interprocedural paths are duplicated, and  dataflow
   analysis  is  performed;  this may lead to sharper data-flow facts along the
   duplicated (hot) paths.

        Preliminary results suggest that this is not a  fruitful  extension  of
   the  original  work.  In the experiments on constant propagation, only small
   increases in the number of constants (on the  order  of  hundreds,  weighted
   dynamically)  are  found.   Meanwhile,  the  transformation  for duplicating
   interprocedural paths incurs a much higher run-time overhead than  the  gain
   made from using the new constants.

        I will present some statistics about our interprocedural path-profiling
   techniques, and discuss directions for future research.


4:00 pm, 2310 CS
OS/Networking Seminar:  Nicholas Coleman, University of Wisconsin, "Matchmaking
   Failure Analysis in Condor"

        The classified advertisement (ClassAd) language has been used to  great
   success  as a matchmaking framework for resource management in Condor.  Pro-
   viders and customers of services submit ClassAds  containing  their  charac-
   teristics,  constraints,  and  preferences  to the Matchmaker.  Ideally each
   customer should be matched with a provider such that  both  principals  meet
   one  anothers  requirements.   However,  whether  due to an incorrectly con-
   structed ClassAd or overly strict constraints (from the provider  or  custo-
   mer), some customers fail to find a match.  At present there is no mechanism
   for providing matchmaking failure diagnostics in Condor.

        In this talk I will present several algorithms for matchmaking  failure
   analysis which identify problematic aspects in a customer's ClassAd.  I will
   discuss two separate (though not exclusive) cases of matchmaking failure:

        1. The customer rejects all available machines

        2. All available machines reject the customer

        Both cases call for breaking down the respective  requirements  expres-
   sions  of  the customer and providers into subexpressions in order to deter-
   mine the cause of failure.  Without loss of generality we assume the expres-
   sions  in  both  cases are in disjunctive normal form (DNF).  In the case of
   the former we seek minimal subsets of conjunctions that always  evaluate  to
   false, as well as maximal subsets that evaluate to true.  In the case of the
   latter we partition provider requirements expressions  by  customer  ClassAd
   attributes  in  order  to  determine which characteristics of a customer are
   unacceptable to a provider.





[← Prev in Thread] Current Thread [Next in Thread→]
  • Today's Events (reminder), David Melski <=