Abstract

Parallel computers come in two varieties: those with shared memory and those without. The former are hard to build; the latter are hard to program. In this paper we propose a hybrid form that combines the best properties of each. The basic idea is to allow programmers to define objects upon-which userdefined operations are performed, in effect, abstract data types. Each object is replicated on each machine that needs it. Reads are done locally, with no network traffic. Writes are done by a reliable broadcast algorithm. A language for parallel programming, Orca, based on distributed shared objects has been designed, implemented, and used for some applications. Its implementation uses the reliable broadcast mechanism. For applications with a high read/write ratio to the shared objects, we show that our approach can frequently achieve close to linear speedup with up to 16 processors. BLURB FOR CENTER COLUMN Parallel computers come in two varieties: those with shared memory an...

Citations