Evaluating the Tera MTA Allan Snavely, Wayne Pfeiffer et al

1
Evaluating the Tera MTAAllan Snavely,
Wayne Pfeiffer et al

• Architectural features
• Massive hardware multithreading
• Flat, randomized memory (no data cache)
• Support for automatic parallelization
• Single programming model for 1 or many processors
• Designed to scale
• Goals of Architecture
• Cover memory and other operational latencies
• Ease burden on programmer
• Exploit multiple levels of parallelism
• Scale
• Goals of SDSC Evaluation
• Funded by NSF to evaluate the MTA for the
  purposes of scientific computing
• Wayne Pfeiffer, Larry Carter PIs

2
Evaluating the Tera MTA Executive Summary

• A few kernels and applications have been found
  for which the MTA achieves higher performance
  than other SDSC machines. Such codes have these
  characteristics
• They do not vectorize well.
• They are difficult to parallelize on conventional
  machines.
• They contain substantial parallelism.
• Examples are codes that involve
• Integer sorting.
• Dynamic, irregular meshes or dynamic, non-uniform
  workloads within a regular mesh.
• Parallel operations (such as a general
  gather/scatter) with poor data locality.
• Single-processor performance of the
  multithreaded Tera MTA (with a 260-MHz clock) is
  typically lower than that of the vector Cray T90
  (with a 440-MHz clock). The T90 is faster than
  the MTA processor for 4 out of 7 kernels and 2
  out of 3 applications compared.
• The MTA processor is appreciably faster for one
  kernel which does an integer sort.
• Single-processor performance of the MTA is
  typically higher than that of cache-based,
  workstation processors. An MTA processor is
  substantially faster than a workstation processor
  for 8 out of 9 applications compared. This
  indicates the effectiveness of multithreading as
  compared to cache utilization.
• Scalability on the MTA is good up to 8 processors
  in many instances and better for kernels than for
  larger applications.
• Very good scalability (parallel efficiency
  between 0.80 and 1.00 on 8 processors) has been
  achieved for 6 out of 7 kernels and 5 out of 11
  applications studied.

3
MTA v.s. IBM Blue Horizon
4
MTA v.s. T90
5
Scalability
6
Symbiosis and Congestion Pricing on MTA

• Allan Snavelys Ph.D. thesis (Fall 2000) Advisor
  Larry Carter.
• Symbiosis A term from Biology meaning the
  living together of distinct organisms in close
  proximity. We adapt that term to refer to an
  increase in throughput and job turnaround that
  can occur when jobs are coscheduled on a
  multithreaded machine.
• Congestion Pricing An area of Economics dealing
  with the right way of pricing a congestion
  externality in such a way that users are caused
  to take cognizance of the impact their usage has
  on others.
• Key Observation Resource sharing among
  coscheduled jobs on a multithreaded machine such
  as the MTA or SMT is very intimate.
• Thesis Jobschedulers which take Symbiosis into
  account, when combined with principles of
  Congestion Pricing, deliver significant
  throughput and turnaround gains and maximize
  global user utility when deployed on
  multithreaded machines.
• See www.sdsc.edu/allans