Clusters of Computational Accelerators

1
Clusters of Computational Accelerators

• Jan Prins
• UNC-Chapel Hill

2
Topics

• Similarity of accelerator architectures
• proof-of-concept kernels for high performance
  applications
• New application areas for accelerator
  architectures

3
Accelerator architectures

• Existing commodity accelerators
• Sony/Toshiba/IBM Cell BE
• Nvidia G80 GPU
• compute-unified device architecture (CUDA)
• ATI R600 (almost)
• Related developments
• Intel demonstrates 80-core TFlop chip
• multicore projects
• March of progress next generation GPUs
• Roadrunner to be based on 2nd gen Cell

4
Cell BE and Nvidia G80
GeForce 8800 GTX
Cell BE
5
Programming for the memory hierarchy
simple parallel memory hierarchy (PMH)
simple uniprocessor memory hierarchy (UMH)
6
Accelerator memory hierarchy
Device Memory
Local Store
Local Store
Vector elts
Vector elts
Vector elts
Vector elts
7
Programming accelerators

• Package inherent parallelism available in problem
• to provide the concurrency and parallel slack
  needed at every level of PMH
• serialize where needed to reach appropriate level
  of reuse
• Programming models
• explicit notion of locality
• CUDA
• UPC

8
Clusters of Accelerators
Scale PMH Peak Perf Cost
Rack Global AddressSpace 20TF 250K
Node Local 400GF 4K
CPU L2/L3
core L1
Accelerator Device 200GF 1K
core Local
SIMD Vector
9
Proof of concept kernels

• Demonstrating performance of accelerator clusters
• challenge is towards the bottom of the parallel
  memory hierarchy
• proof-of-concept kernels can establish viability
  and scaling
• Example
• n-body kernels demonstrated to achieve strong
  performance on Cell and G80
• Consequence
• Folding at home clients developed for Playstation
  and PCs with high-end ATI GPU.
• Full GROMACS acceleration on Cell, NAMD
  acceleration on G80 underway

10
New application domains

• Database and datamining operations
• Stream mining

11
Stream mining applications

• Sampling
• Aggregation
• Summarization
• Clustering
• dimensionality reduction
• PCA, SVD
• subspace clustering
• Classification
• Anomaly Detection

12
Challenges

• Continuous data flow
• Limited storage space
• Limited communication bandwidth through hierarchy
• Detecting and modeling changes
• Visualization

13
Conclusions

• Techniques to effectively exploit accelerator
  clusters are relatively independent of particular
  choice of accelerator
• Application demonstrations can follow spiral
  development model focusing on implementation of
  key kernels
• Data mining and stream mining are important
  application areas that may be well served by
  accelerator architectures