Models of Multiprocessor Computing

Paper: 478
Session: P (talk)
Speaker: Lenoski, Dan, SGI, Mountain View
Keywords: ?

Models of Multiprocessor Computing

Daniel E. Lenoski

Silicon Graphics Computer Systems
Advanced Systems Division


Today's options for high-performance computing include message-passing
MPP systems, clusters of workstations, shared-memory multiprocessors,
and parallel vector processors. Shared-memory multiprocessors provide
the most general-purpose programming model, but generally suffer from
limited scalability. New multiprocessor cache coherence protocols
together with high-performance CMOS VLSI and interconnects are enabling
a new class of machine that scales the shared-memory paradigm.

This talk introduces the scalable shared-memory multiprocessor (SSMP)
and SGI's new Origin servers based on this technology. The Origin
machines are highly modular and permit a single system to grow from a
cost-effective deskside uniprocessor to a supercomputer with hundreds
of processors and hundreds of gigabytes of shared-memory. Nodes within
the system provide a peak of 50 GB/sec of memory bandwidth and are
connected by fat hypercube network that has a peak bisection bandwidth
of 25 GB/sec. These capabilities also enable new levels of I/O
performance including file servers with sustained bandwidths in
excess of 500 MByte/sec and networking interfaces such as SuperHIPPI
that can source and sink data at up to 800 MB/sec simulataneously.

We conclude the talk with a look at the future of multiprocessing.
We predict a merger of the HPC computing models on SSMP and
clusters of SSMP systems.