Efficient shared memory multiprocessing and object-oriented programming

Abstract

Object-oriented techniques are shown to hold promise for addressing the growing speed gap between memory and processors on shared-memory multiprocessors. However, reducing cache misses can be at the expense of more misses from the translation lookaside buffer (TLB -a cache of recent page translations). This paper presents simulation results from use of a library called OOSH (Object-Oriented Library for Shared Memory), in which cache behaviour is measured. Even for a well-optimized application, the library improves speedup by over 20% on 32 processors. On the other hand, timing on a real machine shows that the TLB problem can impact performance by as much as 25%.