Vous pouvez vous abonner à nos annonces de séminaires
                     http://gallium.inria.fr/seminaires/

                             S E M I N A I R E
       __                                  
      / _`  _   /  /  o                       /| /|   __  __  __ __    _  _
     /   ) __) /  /  / /  / /\/|    -----    / |/ |  /  )(_  /  /  ) ) ) __)
    (___/ (_/ (_ (_ / (__/ /   |            /     | (__/ __)(_ (__/ (_/ (_/
                                                
                          I N R I A - Rocquencourt
                     Salle de réunion du bâtiment 16

                         Mardi 14 décembre, 14h30

                               -------------
                               Claudio Russo
                               -------------

                             Microsoft Research

                          ========================
                          A Scalable Joins Library
                          ========================

Polyphonic C\#, C_\omega, the Joins concurrency library and
Concurrent Basic are past MSR Cambridge projects that extend .NET
languages with declarative, high-level message passing constructs,
derived from Fournet and Gonthier's join calculus. The existing
implementations, though sophisticated, are lock-based and the
abstractions come at some cost: programs with high rates of contention
do not scale to large number of processors. We now have a new
implementation of the Joins library that uses ideas from fine-grained
concurrency and non-blocking data structures to achieve much better
scalability. The algorithm is simple and gives promising results. We
validated our scalability results experimentally on declarative,
high-level renditions of six classic coordination problems. Our new
library scales significantly better than the lock based one. In
addition, we scale competitively with bespoke solutions, though we
occasionally suffer from higher constant-time overheads. This talk
will give an overview of the core algorithm; describe the benchmarks
and touch on some optimizations and extensions.  The basic ideas
should be of interest to anyone considering a parallel, shared-memory
implementation of join patterns.

This is joint work with Aaron Turon from Northeastern University.