Supercomputing Applications in Medicine

Phil Andrews, Data Intensive Systems Mgr., Pittsburgh Supercomputing Center.

Talk: http://www.psc.edu/~andrews/talks/sc96.html

Not a "Gimme"

A Supercomputer is not an "ultrafast PC"; an application is not suitable for Supercomputing just because it runs on a computer.

In the future single processor speeds will be relatively invariant, distinguishing feature will be data and communication handling.

When is Supercomputing appropriate ?

Overlap with Classical Supercomputing (e.g., fluids)
Easily Parallelizable (e.g., image comparisons)
Massive Archival Needs (e.g., image databases)

Classical Supercomputing: Heart Modelling

A reasonable match to a vector-based supercomputing. Peskin and McQueen (NYU) used a 16 processor Cray C90 to solve fluid mechanics of blood flow within a heart.

The PSC's "Mario" C90

Parallelizable Applications

Massively parallel machines use large numbers of tightly-couple processors to efficiently handle some problems. Can work with very large amounts of data if problem parallelizes well.

Building "signatures", or comparing raw images fits well on an MPP machine when each processor can hold a single image. Requires excellent data flow.

The PSC's "Jaromir" T3E

Archival:

Supercomputing centers build ancillary services such as a very large, rapidly accessible archival and excellent network access. Medical image databases are very demanding, as compression is rarely satisfactory.

The IBM 3494 ATL at PSC has 8 IBM Magstar drives, delivering >10 MB/s each from 10 GB tapes (23 GTB after compression). Capacity is 2400 tapes.

The Storage Tek silos hold 6,000 tapes each, we plan to retrofit MagStar drives in Jan '97. Capacity should be several hundred TB. Hope to use parallel I/O to deliver large files at >80 MB/s from tape.