MegaComputing A Reflection on the Challenges

1
MegaComputingA Reflection on the Challenges
Presentation to SC00 Panel on Megacomputing

• Thomas Sterling
• California Institute of Technology
• And
• NASA Jet Propulsion Laboratory
• November 10, 2000

2
Performance Development
60G - 400 M4.9 Tflop/s 55Gflop/s, Schwab 15,
1/2 each year, 209 gt 100 Gf, faster than Moores
law,
3
Requirements for High End Systems

• Bulk capabilities
• performance
• storage capacities
• throughput/bandwidth
• cost, power, complexity
• Efficiency
• overhead
• latency
• contention
• starvation/parallelism
• Usability
• generality
• programmability
• Reliability
• Confidence
• Security
• Correctness
• accessibility

• Controlled fusion for power propulsion
• Molecular modeling protein folding
• Drug design and genomic mapping
• Aerospace interdisciplinary design
• Human anatomy modeling e.g. heart
• Climate modeling weather prediction
• Automated Innovation through optimization

4
Challenges

• Degree of problem parallelism
• Usable granularity of atomic work module
• Data I/O capacity requirements per work module
• Latency sensitivity of computation
• Failure tolerance and automated checking/rollback
• Security
• data and computation results (client
  confidentiality)
• resident information (server confidentiality)
• returned results (protection of client)
• input data and code (protection of server)
• Business model
• Is free really free
• Is there enough of the right kind of workload
• Transferring cost from iron to fiber
• System administration
• Are the majority of data intensive problems
  really possible customers

5
Epochs of Apex Computing

• ALU speed and Memory capacity Constrained
• Components too expensive
• Solution
• VLSI
• Latency Constrained
• Distance, Overhead, Contention
• Solution
• Vector pipelining
• Cacheing
• Multithreading
• Network Bandwidth
• Locality management through partitioning
• Trust Constrained
• Collaborative computing

6
Were We Wrong?

• Latency doesnt matter to efficiency of
  computation
• Bandwidth is cheap enough to send everything all
  around the country/world
• Its easier to administer a nation straddling
  megacomputer than a Beowulf
• Fault tolerance is guaranteed in Grid-land
  although difficult in a single computer
• The challenges of programming a parallel computer
  are much harder than programming a continent
  sized distributed computer

7
My guess is

• Its different from Grids although may use much of
  the same software technology
• Its going to work for the right kind problems
• Very large
• Wonderfully Parallel
• Response-time flexible
• Security insensitive
• It will appeal to certain class of users and
  suppliers
• Its cool (seti _at_ home effect)
• Provides access to OOM burst peak requirements
• It will produce some extraordinary results
• It may be the genesis of P-1

8
A Final Concern

• Will it be an enabler
• Provide unprecedented computing capacity
• Another OOM advance in price-performance
• Will it be an inhibiter
• Only suitable problems will be pursued
• The demise of innovative computer architecture
• Will efficiency continue to degrade
• Will the internet compute capacity continue to
  explode
• What happens when everyone has a computer
• Will most of our processing go off shore
• What will P-1 do when it wakes up