Erik P. DeBenedictis

1
Sandia Petaflops WorkshopOverview for SOS7

• Erik P. DeBenedictis

Sandia is a multiprogram laboratory operated by
Sandia Corporation, a Lockheed Martin
Company,for the United States Department of
Energy under contract DE-AC04-94AL85000.
2
Fishing

• I was a kid in Alaska during 5th grade
• Bought Ted Williams (?) fishing kit for 10
• For a year threw hook in water, waited 5 minutes
  and pulled it out
• Never caught anything
• Conclusion Fishing mostly an exercise in
  patience
• No fishing for me

3
Objective

• What do you put into a supercomputer chip in year
  20XX
• Consumer microprocessor (Hammer, Power, Intel, )
• PIM
• Whose?
• FPGA
• Why?

4
Domain of Optimization

• Sandia/DOE Emphasis on Supercomputer Architecture
• Interests
• Architectures that run our applications well
• Scalability over the long haul 10 years
• Competing approaches
• Preserve code base for a particular CPU line
  (X86)
• Repurpose commercial hardware
• Peak FLOPS

5
Value Chain
TechnologyDiegertDeBenedictis
ArchitectureBurgerSterlingTomkins
Applications HoisieChristopher Bender
Systems Support Maccabe
ValueDelivery via batch LeungRudolph Tomkins
6
Talks 1 2

• Carl Diegert
• Physical packaging topic
• To minimize delay, shrink machine
• This doesnt cut power and eventually you reach
  cooling limits

• Erik DeBenedictis
• Technology scaling and balance factors topic
• Moores Law permits oracle into the future
• Project future of balanced supercomputers

7
SIA Semiconductor Roadmap

• Generalization of Moores Law
• Projects many parameters
• Years through 2016
• Includes justification
• Panel of experts
• known to be wrong
• Size between Albuquerque white and yellow pages

8
Talks 3 4

• Thomas Sterling
• Gilgamesh/MIND PIM Architecture
• A physically realizable supercomputer candidate
• Unconventional programming

• Doug Burger
• TRIPS PIM Architecture
• A physically realizable supercomputer candidate
• Unconventional programming

9
TRIPS
10
Gilgamesh
11
Talks 5 6

• Jim Tomkins
• Traditional Roadmap of DOE Supercomputer
  purchases
• Programmable
• Physically realizable?

• Adolfy Hoisie
• (Giving Kerbysons talk, who had family
  emergency)
• Estimating performance of key Government
  applications on machines that have not been build
  yet

12
Talks 7 8

• Thomas Christopher
• Fusing Moores Laws clairvoyance with
  application performance estimation
• Predict performance of machines based on abstract
  trends 10 years into the future

• Barney Maccabe
• FAST-OS
• Can we put a runtime system on whatever
  transpires?

13
Estimating Performance
14
Talks 9 10

• Michael Bender
• Are there ways we can incrementally change the
  way we program to meet the limitations of the
  laws of physics?

• Vitus Leung
• Fast programs are only part of the picture the
  batch scheduler runs the show
• Batch schedulers typically waste 15 40 of a
  machine
• What new issues will peta-scale supercomputers
  introduce?

15
Talk 11

• Larry Rudolph
• In detail, how do we value a peta-scale
  supercomputer
• Peak flops?
• Throughput?
• Response time?

16

• 48 GFLOPS Hammer ? 20 picoseconds/flop
• 2 FLOPS/clock ? 40 ps clock period
• Speed of light 1 foot/ns
• Or .04/40 ps
• On a chip say .5 c
• .02 per clock cycle
• 50 clock cycles to get across 1 chip
• Challenge
• Enormous pipelining