Runtime Locality Optimizations of Distributed Java Applications

  • Tagung:

    Konferenzartikel 

  • Autoren:

    Christian Hütter
    Thomas Moschny 

  • Summary

    In distributed Java environments, locality of objectsand threads is crucial for the performance of parallelapplications. We introduce dynamic localityoptimizations in the context of JavaParty, aprogramming and runtime environment for parallelJava applications. Until now, an optimal distributionof the individual objects of an application has to befound manually, which has several drawbacks.

    Based on a former static approach, we develop adynamic methodology for automatic localityoptimizations. By measuring processing andcommunication times of remote method calls atruntime, a placement strategy can be computed thatmaps each object of the distributed system to itsoptimal virtual machine. Objects then are migratedbetween the processing nodes in order to realize thisplacement strategy. We evaluate our approach bycomparing the performance of two benchmarkapplications with manually distributed versions. It isshown that our approach is particularly suitable fordynamic applications where the optimal objectdistribution varies at runtime.

  • Jahr:

    2008 

Beteiligte Mitarbeiter (zufällige Reihenfolge)
Titel Vorname Nachname

Bibtex

@inproceedings{,
author={Christian H{\"u}tter, Thomas Moschny},
title={Runtime Locality Optimizations of Distributed Java Applications},
year=2008,
month=Feb,
booktitle={Proceedings of the 16th Euromicro International Conference on Parallel, Distributed and Network-Based Processing (PDP 2008), February 13-15, 2008, Toulouse},
publisher={IEEE Computer Society},
abstract={In distributed Java environments, locality of objectsand threads is crucial for the performance of parallelapplications. We introduce dynamic localityoptimizations in the context of JavaParty, aprogramming and runtime environment for parallelJava applications. Until now, an optimal distributionof the individual objects of an application has to befound manually, which has several drawbacks.Based on a former static approach, we develop adynamic methodology for automatic localityoptimizations. By measuring processing andcommunication times of remote method calls atruntime, a placement strategy can be computed thatmaps each object of the distributed system to itsoptimal virtual machine. Objects then are migratedbetween the processing nodes in order to realize thisplacement strategy. We evaluate our approach bycomparing the performance of two benchmarkapplications with manually distributed versions. It isshown that our approach is particularly suitable fordynamic applications where the optimal objectdistribution varies at runtime.},
pages={149-156},