EECS Publication
HPCS Library Study Effort
Jack Dongarra, James Demmel, Parry Husbands, and Piotr Luszczek
In this report we present our research into the implementation of numerical libraries using the proposed HPCS languages. Faced with the fact that the community has very little application experience (the implementations are not yet mature) with these languages, we chose a somewhat atypical approach: perform a case study of parallel LU factorization and determine how this kernel can be implemented in the languages. As such we decided to gather various algorithmic techniques that have been successful and make connections to specific HPCS language features. We settled on parallel LU factorization for a variety of reasons: - It is a well known, understandable kernel - Many implementations exist that span the performance spectrum - Getting it to perform well in parallel on distributed memory machines reveals many programming issues, solutions to which aren't well represented in traditional languages. - In Section 2 we give a short description of the algorithm and outline some of the roadblocks to high performance. Section 3 presents some of the abstraction issues that arise when comparing the implementation of different versions of the algorithm in different languages. Section 4 contains our survey of the implementations. We detail our observations regarding implementing a high performance LU code in an HPCS language in Section 5 and conclude in Section 6.
Published 2008-05-27 04:00:00 as ut-cs-08-617 (ID:92)