1994 MIT Student Workshop on Scalable Computing

MPPs are Unreliable

• For a 64+4 node KSR running for 1.38 years, MTTI was 1.76 days.
• For a 56-compute-node Paragon, over a period of 90 days, MTTI was 1.96 days.
• For a CM-5 running for a year, there were ~700 time-sharing demo crashes (2 per day). There were 10 reboots of Scout in less than thirty days.
• In the MIT Computer Architecture Group's set of 40-50 workstations, about 1 goes down a month. A central server fails once every 100-200 days.
• Supercomputer companies are, perhaps, as unrealiable as their machines (e.g., KSR, TMC, ...). [Note, this was written before those companies' recent announcements.]

Workstation networks are fast (and getting faster)

• This has been shown by Thekketh's research.
• Consider the following read latencies for various systems:

  					40 bytes 	80 bytes
  	Decstation 5000/200 ATM 	37		55
  	Intel Delta			71		75
  	CM-5				91		96
  	CM-5 with Active Messages	28		33
  

• Consider the throughput for a massive-write:

  	Throughput (massive write)
  	Decstation		4.3 Mb/Sec
  	Delta			10  Mb/Sec
  	CM-5			8.3 Mb/Sec
  

• Consider the ATOMIC work: 1.28Gb/s per link. With all equipment included, a link is about $1200.

Case Study: The University of Arizona put C* on a network of workstations

• 8 50Mhz HP Snakes
• Performance was, on average, within a factor of two of 32-node CM-5 with 4 vector units per node.
• Won when more communication than computation (because of vector units).

What does a super have that workstations don't?

• Huge price tag
• More sys admins
• Pretty lights