The Once and Future SciDAC

1
The Once and Future SciDAC
with apologies to T. H. White

• Thom H. Dunning, Jr.
• National Center for Supercomputing Applications
• and Department of Chemistry
• University of Illinois at Urbana-Champaign

National Center for Supercomputing Applications
University of Illinois at Urbana-Champaign
2
SciDAC The Program
Advances in the simulation of complex scientific
and engineering systems provide an unparalleled
opportunity for solving major problems that face
the nation in the 21st Century.
3
SciDAC Goals
Create a Scientific Computing Software
Infrastructure that bridges the gap between
applied mathematics computer science and
computational science in the physical, chemical,
biological, and environmental sciences

• Scientific Application Codes
• Develop mathematical models, computational
  methods, and scientific codes to take full
  advantage of the capabilities of terascale
  computers
• Computing Systems and Mathematical Software
• Develop software infrastructure to accelerate the
  development of scientific codes, achieve maximum
  efficiency on high-end computers, and enable a
  broad range scientists to use simulation in their
  research
• Collaboratory Software
• Develop network technologies and collaboration
  tools to link geographically separated
  researchers, to facilitate movement of large
  (petabyte) data sets, and to ensure that academic
  scientists can fully participate in these
  activities

4
SciDAC Goals II
Create a Scientific Computing Hardware
Infrastructure that is robust, agile, and
flexible

• Flagship Computing Facility
• To provide computing resources to address a broad
  range of scientific problems
• Topical Computing Facilities
• To ensure that the most effective and efficient
  resources are used to solve each class of
  problems
• Experimental Computing Facilities
• To guide advances in computer technology to
  ensure that scientific computing has the
  resources that it needs in the future
• ESNet
• To support research in a connected world

5
SciDAC Circa 2001
COLLABORATORIES
COMPUTING SYSTEMS SOFTWARE
M A T H E M A T I C S
O P E R A T I N G
D A T A G R I D S
S Y S T E M
Data Analysis Visualization
Programming Environments
Scientific Data Management
Problem-solving Environments
BES, BER FES, HENP
6
SciDAC Score Card
Goal Status Comments
Scientific Challenge Codes Excellent progress in selected areas, but many areas poorly supported or even neglected
Computing Math Software Excellent progress, but some areas need additional support
Collaboratory Software Good progress, but little used
Flagship Computing Facility Two facilities established, NERSC and NLCF, but
Topical Computing Facilities QCDOC and MSCF, but many opportunities still unexplored
Experimental Computing Facilities Little progress
7
After 5 YearsIs SciDAC Still Needed?

• Yes!

8
After 5 Years Does SciDACNeed More Funding?

• Yes!

9
Central Dogma
The central dogma of SciDAC is the close coupling
between computer hardware and computer software
Hardware
SciDAC Multidisplinary Teams
Porting Revision Rewriting
Enhanced Performance (can be dramatic)
Software
Changes in computer hardware requires changes,
often major changes, in computer software.
Responding to such changes in a timely manner
requires a multidisciplinary approach.
10
The Coming Revolutionin Computing

• The Free Lunch Is Over A Fundamental Turn
  Toward Concurrency in Software
• Herb Sutter in Dr. Dobbs Journal 30(3), March
  2005

11
The GHz Race
At the 2000 IEEE International Electron Devices
Meeting, Intel announced that it expected to
produce a 10 GHz microprocessor by 2005. The
fastest Intel microprocessor today runs at 3.8
GHz (Intel Pentium 4). It was introduced six
months ago. At its presentation of the 6XX series
of Prescott, Intel stated that it is committed to
adding value beyond GHz.
12
Increasing Computer Performance

• Increasing Clock Frequency
• Pentium 60 MHz to 3,800 MHz in 12 years
• Resulted in 80 of performance increase

13
The Heat Problem
Rocket Nozzle
1000
Nuclear Reactor
Pentium 4 (Prescott)
100
Pentium 4 (Willamette)
Watts/cm2
Pentium III
Hot Plate
Pentium II
10
Pentium Pro
Pentium
i386
i486
1
1.5?
1.0?
0.7?
0.5?
0.35?
0.25?
0.18?
0.13?
0.1?
0.07?
Increasing Frequency
Courtesy of Bob Colwell
14
Managing the Heat Load
Liquid cooling system in Apple G5s
Heat sinks in 6XX series Pentium 4s
15
Leakage CurrentFrom Minor Nuisance to Chip Killer
Dissipated Power CV2f
Power (W)
70
130
90
250
180
Process Technology (nm)
16
Means of Increasing Performance

• Increasing Clock Frequency
• From 60 MHz to 3,800 MHz in 12 years
• Has resulted in 80 of performance increase
• Execution Optimization
• More powerful instructions
• Execution optimization (pipelining, branch
  prediction, execution of multiple instructions,
  reordering instruction stream, etc.)

17
Microarchitecture Trends
106
Multi-Threaded, Multi-Core
Pentium 4 and Xeon Architecture with
HT Multi-Threaded
105
Era of Thread Parallelism
Pentium 4 Architecture Trace Cache
104
MIPS
Pentium Pro Architecture Speculative Out-of-Order
103
Era of Instruction Parallelism
Pentium Architecture Super Scalar
102
101
1980
1985
1990
1995
2000
2005
2010
Adapted from Johan De Gelas, Quest for More
Processing Power, AnandTech, Feb. 8, 2005.
18
Means of Increasing Performance

• Increasing Clock Frequency
• From 60 MHz to 3,800 MHz in 12 years
• Has resulted in 80 of performance increase
• Execution Optimization
• More powerful instructions
• Execution optimization (pipelining, branch
  prediction, execution of multiple instructions,
  reordering instruction stream, etc.)
• Larger Caches
• On-chip caches to ameliorate the growing
  disparity between processor speed and memory
  latency and bandwidth

19
Moores Law Still Holds
10
11
4G
2G
10
10
1G
512M
Memory
256M
10
9
128M
Itanium
Microprocessor
64M
10
8
Pentium 4
16M
Pentium III
10
4M
7
Pentium II
1M
10
6
Pentium
256K
Transistors Per Die
i486
64K
10
5
i386
16K
80286
4K
10
4
8080
1K
8086
10
3
4004
10
2
10
1
10
0

60

65

70

75

80

85

90

95

00

05

10
Source Intel
20
Increasing Caches Montecito
21
Means of Increasing Performance

• Increasing Clock Frequency
• From 60 MHz to 3,800 MHz in 12 years
• Has resulted in 80 of performance increase
• Execution Optimization
• More powerful instructions
• Execution optimization (pipelining, branch
  prediction, execution of multiple instructions,
  reordering instruction stream, etc.)
• Larger Caches
• On-chip caches will continue to increase in size
  and help mitigate disparities in computer
  subsystem performance

22
New Technologies for Computers

• Low power processors

23
IBM Blue Gene Systems

• LLNL BG/L
• 360 teraflops
• 64 racks
• 65,536 nodes
• 131,072 processors
• Node
• Two 2.8 Gflops processors
• System-on-a-Chip design
• 700 MHz
• Two fused multiply-adds per cycle
• Up to 512 Mbytes of memory
• 27 Watts

24
Technologies for Petascale Computers

• Low Power Processors
• Need unprecedented application software
  scalability
• Application codes must scale to 100,000s of
  processors
• Need ability to recover from continual processor
  loss

25
New Technologies for Computers

• Low Power Processors
• Need unprecedented scalability
• Application codes must scale to 100,000s of
  processors
• Need ability to recover from processor loss
• Multicore Chips

26
Architecture of Dual-Core Chips

• IBM Power5
• Shared 1.92 Mbyte L2 cache
• AMD Opteron
• Separate 1 Mbyte L2 caches
• CPU0 and CPU1 communicate through the SRQ
• Intel Pentium 4
• Glued two processors together

27
Intel Processor Roadmap
28
New Technologies for Computers

• Low Power Processors
• Need unprecedented scalability
• Application codes must scale to 100,000s of
  processors
• Need ability to recover from processor loss
• Multicore chips
• Need to better understand a number of
  architectural issues
• Memory bandwidth
• Cache contention

29
Other Promising Technologies

• Field Programmable Gate Arrays (FPGAs)
• Capabilities increasing rapidly (riding silicon
  technology curve)
• Need efficient software development tools
• Heterogeneous Computer Systems
• Different types of processors in single system
• Vector processors, superscalar processors, FPGAs
• High speed interconnect linking all processors
• May be especially advantageous for some
  applications, e.g., multiphysics applications
• Many Other New Ideas
• DARPA High Productivity Computing System program
• Universities Sterling, Dally,

30
SciDACPathway to the Future

• Questions?