Next: Overhead-Based Performance Estimates
Up: Algorithm Performance on
Previous: Algorithm Performance on
We believe that a power utility's interests in future parallel
architectures will be in scalable parallel processors (SPPs) rather
than massively parallel processors (MPPs), because:
- the compatibility of SPP nodes with networked desktop
computing resources contributes to reduced business overhead
costs,
- small to midsized SPPs offer an improved cost/performance
ratio when compared to small MPPs.
We can expect future SPP architectures to be similar to the IBM
SP-series with 2-64 processors interconnected by a non-blocking,
high-bandwidth, switched network [16]. Internode
communications performance may soon approach that of the Cray T3D
massively parallel computer (1
second latency and 300 megabytes
per second bandwidth) [43]. When comparing the single
processor performance of the CM-5 (a 33 MHz Sparc microprocessor from
Sun Microsystems) [6] with a node of the Cornell
Theory Center SP1 or Northeast Parallel Architectures Center (NPAC)
SP2 (a 62.5 MHz IBM RS/6000 model 370 four command superscalar
microprocessor), we have shown in section 7.1, that
the IBM RS/6000 microprocessor in the SP1 is 6.6 times faster than the
33 MHz Sparc microprocessor when comparing empirical data from our
algorithm run on a single processor. The speed for the microprocessor
in the Cornell Theory Center SP2 is even 50% faster than the SP1
RS/6000 microprocessor. In the near-future, it will be feasible to
get four times the individual processor power that is now available on
the SP1, so it is conceivable that the future generation of SPP
microprocessors will be 25 times as fast as those used in the Thinking
Machines CM-5. Some of this processing power may come from placing
multiple shared-memory processors per SPP node
[16],
If SPP node processor capability increases by a factor of 25 relative
to the Thinking Machines CM-5, communications capabilities must
improve by at least as much if parallel sparse direct linear solver
performance for power systems applications is to have equal or better
multiple processor speedup. In other words, the
computation-to-communications ratio for the SPP must remain constant
or improve in order that SPP speedup remains constant or improves.
Next: Overhead-Based Performance Estimates
Up: Algorithm Performance on
Previous: Algorithm Performance on
David P. Koester
Sun Oct 22 17:27:14 EDT 1995