This three talk series is organized as a mini-course in modern parallel computer architecture design.  The talks focus on fundamental issues as opposed to case studies but are backed up by some case study snippets which are used to illustrate the basic design alternatives.  The series starts with a tutorial on modern parallel architecture styles.  The second talk will integrate these issues for modern convergence machine designs. The third talk will focus on the design issues of the communication fabric.

Modern microprocessors provide such a high level of performance per cost that most of today's vendors and researchers choose to use them rather than design their own.  The problem then becomes how to integrate the memory and communication resources in an efficient manner.  The parallel programming community has begun to standardize on two distinct models of computation,  namely message passing and distributed shared memory.  This talk will survey the sets of design options and performance issues that must be
addressed in the creation of a parallel machine.  Topics will include latency and bandwidth concerns in supporting message passing; protocol, coherence, and occupancy issues is shared memory systems; system bus design, I/O architecture, and network interface issues.

Given the predominance of message passing and shared memory programming models, the obvious challenge is to support both in a cost effective manner.  Researchers at Stanford (Flash), Utah (Avalanche), Wisconsin (Typhoon), and Princeton (Shrimp) are meeting this challenge in a variety of ways.  This talk will focus on the basic design issues and show examples of solutions from these four specific efforts.  These approaches will be contrasted with the current commercial approaches by IBM (SP-2) and SGI (Origin 2000).

Clearly parallel system performance depends critically on the speed of the underlying communications fabric.  This talk will start with a tutorial on the key points in the design space and the inherent trade-offs which must be made in choosing a particular design decision.  The talk will conclude with a comparative analysis of 5 particular fabric designs: Chaos (U. Washington), Post Office (HP Labs), Myrinet (Myricom), Spider (SGI/Cray), and Vulcan (IBM).
 
 Return to the main page