Workshop on experiences with building distributed and multiprocessor systems
Chuck Koelbel, Eugene Howard Spafford, George Leach · ACM SIGOPS Operating Systems Review · 1990
Many aspects of parallel software were discussed at the first USENIX/SERC Workshop on Distributed and Multiprocessor Systems, held October 5-6 in Fort Lauderdale, Florida.The workshop was sponsored by the USENIX Association and the Purdue/Florida Software Engineering Research Center (SERC).The IEEE-CS Technical Committees on Operating Systems (TCOS) and Distributed Computing (TCDC) and the ACM Special Interest Groups on Operating Systems (SIGOPS) and Software Engineering (SIGSOFT) were sponsors-incooperation.Gene Spafford of Purdue University was program chair and George Leach of AT&T Paradyne served as general chair.At the workshop, researchers from industry and academia presented 25 papers in eight sessions.The papers were selected through a refereeing process from over 50 submissions.In addition, a Work in Progress session featured six short reports and an informal session was organized to discuss topics raised in one of the early sessions.At the "Objects and Virtual Memory" session, Franz Kaashoek of Vrije Universiteit, Yousef Khalidi, formerly of Georgia Tech, and Gary Johnston of the University of Illinois reported on three research projects implementing shared virtual memory in a distributed environment.Their approaches differed in details, but the basic thrust of all three projects was to encapsulate data in objects and use those objects as the basic unit to be shared among machines.Although nobody disputed the results of these projects, there was a significant group of attendees who claimed that distributed shared memory was the wrong way to go.At a specially-organized rebuttal session that night (entitled "Distributed Memory Is Easy ... but Wrong"), those critics argued that such systems hid the cost of data sharing from the user, thus either encouraging inefficient programming or making performance tuning impossible.Several alternatives were offered, many of which were included in other workshop papers.The point was made by several speakers that the model of programming in a system should be tied closely to both the system architecture and to the application under consideration.By that standard, distributed memory was considered a definite advantage for some systems but less valuable for others.