1. Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
An Easy Path to Exascale
Scott Pakin
Applied Computer Science Group
Los Alamos National Laboratory
25 April 2012
Slide 1
2. Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
Disclaimer
The opinions expressed herein definitely do not reflect the positions of
Los Alamos National Laboratory, the National Nuclear Security
Administration, or the United States Department of Energy
Slide 2
Please don’t cut off my funding
3. Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
Background
Many of us have been to lots of
exascale meetings
If these meetings forbade the
use of the words
• energy
• resilience
• programmability
• co-design
the meetings would be a lot
shorter
Slide 3
(but perhaps more fun)
4. Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
Exascale Refrains
We don’t know how to reduce
system power
We don’t have a good way to
tolerate frequent faults
As component heterogeneity
increases, programmability
decreases
Slide 4
Exascale is
hard!
5. Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
Solution
How can we deal with all of the
problems of exascale?
Slide 5
6. Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
My Proposal
Slide 6
What if instead of building an exascale system, we gave each of 1000
researchers his/her own, private, petascale system?
exascale
petascale
7. Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
My Proposal
Power problem solved
Resiliency problem solved
Programmability problem solved
Increased productivity
Lower cost
Slide 7
What if instead of building an exascale system, we gave each of 1000
researchers his/her own, private, petascale system?
Different locations → distribute
load across the power grid
Fewer components → faults
occur less frequently
Less complex hardware → can
stick with MPI+X (where X=∅)
Dedicated systems → no long
queues waiting for job to start
More aligned with vendor
roadmaps → no cajoling needed
8. Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
What About Applications that Require Exascale?
Approaching a singularity where we don’t have any applications
(except perhaps LINPACK)
Slide 8
Numberofsuitableapps.
Sequential
Small
SMP
Large
NUMA
Distributed
memory
Heterogeneous
architecture
Exascale
(power, resiliency)
Zettascale
(quantum
decoherence?)
Yottascale
(inter-D-brane
interactions on
11-dimensional
M-theoretic
superstrings?)
9. Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
We May be Facing a Petascale Performance Asymptote
Parallelism
• By Amdahl’s Law, if your application is less than 99.9% parallelized, you’ll see at
best only petascale performance on an exascale system
Resilience
• Performing the same work three times gives you 1/3 of the remaining performance
Power
• To stay within the power budget, some components may have to be turned off or
downclocked, lowering performance even further
1E+0
1E+2
1E+4
1E+6
1E+8
1E+10
1E+12
1E+14
1E+16
1E+3 1E+6 1E+9 1E+12 1E+15 1E+18 1E+21 1E+24
Theoretical peak performance
Application
performance
Mega Giga Tera Peta Exa Zetta Yotta
10. Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
Summary
Exascale computing poses numerous challenges
• Energy, resilience, programmability
Relatively few applications both require exascale and can be
programmed to exploit complex exascale hardware
Many applications can take advantage of petascale
• Limited by access to machines
• (Long queue if you want the machine for yourself)
Proposition: Buy 1000 research teams their own petascale computer
• Increases scientists’ productivity
• Why put ourselves through the pain of an exascale before its time?
• I’ll wait until I can borrow an exascale cell phone from my grandkids
It’s not like there isn’t still exciting research to be done at petascale
Slide 10