Transcript Some Thoughts on Technology and Strategies for Petaflops

Some Thoughts on
Technology and Strategies for
Petaflops
Possible paths to Petaflops
• Traditional Commodity Clusters
• Leverage Moore’s law on GP microprocessors
• Interconnect and memory bandwidth problems
• Type C machines
• DARPA HPCS paths (e.g. Cascade etc.)
• Embedded systems based Clusters
• QCDOC one example
• BG/L another example
Rick Stevens
Argonne  Chicago
Beyond Commodity Clusters
• Improved design capability
• Small groups can design SoCs
• Small groups can gain access to state of the art
fabrication capabilities
• Design cycles are getting shorter thanks to
increasing availability of off-the-shelf IP
• Blue Logic, MIPS, etc.
• QCDOC example
Rick Stevens
Argonne  Chicago
Rick Stevens
Argonne  Chicago
Hardware/Software Co-design
• Application kernels
• Simple “FORTRAN” like C code - well
behaved basic blocks with performance
requirement annotations
• Compiler builds performance model for each basic
block
• Decision point based on performance estimate
• Compile for GPU or synthesize logic/FGPA code
• Generate glue code/runtime
Rick Stevens
Argonne  Chicago
Special purpose SoCs
• Networking Processing Units
• Core of fast IP switches and routers
• Many companies producing 10Gbps components and
moving towards 40 Gbps parts
• DSPs
• Cell phone base stations.. Signal processing and array on
a chip processors
• Example is 2 GHz, 175 Million transistors 64 processor
DSP array, several hundred dollars a chip in quantities of
1,000.
Rick Stevens
Argonne  Chicago
Graphics Accelerators
• NVIDIA Geforce4 example
• > 100 M transistors
• High-speed (QDR) RAM interface > 10 GBps
• Moving towards General purpose processors
• Cg programming language (programmable shaders)
• Evolving to become faster than the main CPU on
a commodity based node
• Pentium or Itanium2 process becomes a service
processor?
Rick Stevens
Argonne  Chicago
Extendable Cores
• Possible target for HPC Hardware/Software Codesign
• Provides a reconfigurable node platform
• Xilinx virtex-pro
• Multiple PowerPC cores (1-4)
• Millions of gates of FPGA
• Clock rates lag high-performance chips
• Other vendors producing similar things
• MIPS cores, SPARClite cores, etc.
Rick Stevens
Argonne  Chicago
Billion Transistor Dies by 2005/6
• Design challenges and opportunities
• Many 32 bit cores available < 500,000 transistors
• Several 64 bit cores available < 2,000,000
transistors
• Complete SoC libraries becoming available (e.g.
Blue Logic, etc.)
• Unprecedented opportunity for semi-custom
node architectures based on SoC technologies
Rick Stevens
Argonne  Chicago
Design Tools are Improving
• We can start to think in terms similar to desktop
publishing from 20 years ago
• Mass customization will become possible but:
• What design Macros are needed ?
• How to involve algorithms and applications developers in
the design process ?
• How to connect with systems software (OS, runtime,
libraries)?
Rick Stevens
Argonne  Chicago
Evolution of Commodity Clusters
I/O
Commodity Network
GPU/Node
…..
SoCs
GPU/Node
SoCs
…..
Rick Stevens
High-Performance Interconnect
O(1000) nodes
GP services
O(100K) nodes
Semi-custom or
Reconfigurable
Argonne  Chicago
Systems Software for SoCs
• Embedded Processor Systems Software
• DSP: real-time OS/Runtime ~40K on chip
FLASH ROM (shadow RAM), off chip
extensions for future
• NPUs: real-time runtime support < 100K
typically, some general purpose co-processors
(Linux typically used in Juniper)
• Graphics processors on chip runtime support
upgradeable via device drivers
Rick Stevens
Argonne  Chicago
A Few Recommendations
• Comprehensive applications studies
• To determine feasibility of acceleration via semi-custom
SoC/CLoCs
• To understand what OS functions are actually required
for full HPC applications
• Establish some design challenges
• Pick several core algorithms (besides lattice gauge) and do
some paper designs to validate the possible advantages of
SoC based approaches
• An augmented cluster testbed
• GP Linux cluster with SoC/CLoC based compute
backends
Rick Stevens
Argonne  Chicago