IMMD-IV UP HELP Feb 13, 2004

Using Locality Information in Userlevel Scheduling

Technical Report TR-I4-95-14

Using Locality Information in Userlevel Scheduling
Martin Steckermeier, Frank Bellosa
english April 1996, 41 pages
TR-I4-95-14
[Abstract] [Full Paper (ps,http) , 116 kB] [Full Paper (pdf) , 296 kB]

Abstract: In the past few years, MIMD parallel computers have become important not only in the field of high performance scientific computing, but also as ordinary compute servers. Applications that can not be parallelized by an appropriate compiler get more and more parallelized by use of the threads programming model. Especially for machines having large caches and/or non uniform memory access special care has to be taken for an efficient handling of applications with a huge number of threads. Locality of data has to be taken into consideration as well as the memory access behaviour of threads to assure high cache reusage. This paper examines several techniques that are necessary for efficient thread management on userlevel. Algorithms and mechanisms for scheduling as well as for synchronization are analysed for their suitability in userlevel thread libraries and the importance of using locality information is pointed out. Measurements with a prototype show the superiority of these concepts.


Generated by tr2html V1.3 - © 1998-1999 Michael Golm.