|
|
| Line 59: |
Line 59: |
| | | | |
| | ==Papers== | | ==Papers== |
| − | ===Landscape of parallel computing research: a view from berkeley===
| |
| − | ----
| |
| − | * good intro paper.
| |
| − | * 3 sections
| |
| − | ** software
| |
| − | ** hardware
| |
| − | ** performance/programming models
| |
| − | ** conclusion
| |
| − | * probably cover 1st section w/ dwarfs, plus conclusion, although everything is good
| |
| − | * General introduction to world of // computing in 2006
| |
| − | * old wisdom/new wisdom ==> '''world of // computing is changing'''
| |
| − | * moore's law
| |
| − | :::"Key to this approach is a layer of libraries and programming frameworks centered on the 13 computational bottlenecks ("dwarfs") that we identified in the original Berkeley View report. "
| |
| − | * '''A dwarf is a class of algorithms that tend to have a common pattern of communication and computation'''. Dwarfs are well-defined targets from algorithmic, software, and architecture stand point.
| |
| − | * Page 3: multicore is not the answer. Manycore is.
| |
| − | * High Performance Computing has good ideas and good solutions that map well to current parallel solutions==>use their solutions!
| |
| − | * Page 5: '''We have a break in progress in computing: new directions should be given'''
| |
| − | * Physics becomes important as size gets lower. Power is important.
| |
| − | * latency and bandwidth: '''transfer faster, wait longer'''.
| |
| − | * Page 6: compilers are '''a bout de souffle'''
| |
| − | * 20 more years of Moore's law only? 70 years?
| |
| − | * Page 6: '''Any speed up via parallelism is a success'''
| |
| − | * Page 7: we have been concentrating on massively parallel architectures, but this may apply to new forms of parallelism
| |
| − | * Page 7: good match between scientific // approaches and current parallel (non scientific) domains
| |
| − | [[Image:communicationPatterns.png|right]]
| |
| − | * Page 7: 7 Dwarfs
| |
| − | ** ==> question: what's a dwarf? (google berkeley + dwarf)
| |
| − | ** Dwarfs replace benchmarking in a way
| |
| − | ** high level of abstraction
| |
| − | ** can be used to evaluate new prototypes
| |
| − | * '''The word is a new unit of information'''
| |
| − | * Page 11: 3 new areas for adoption as dwarf?
| |
| − | ** machine learning
| |
| − | ** database software
| |
| − | *** Page 12: map-reduce
| |
| − | ** computer graphics
| |
| − | * Page 14: introduces 6 new dwarfs
| |
| − | * Intel has similar breakdown of paradigms that it calls RMS: Recognition, Mining, and Synthesis.
| |
| − | * '''A parallel to be made: IBM 360 and microprocessors. Now multicore can benefit from High Performance Computing'''
| |
| − | * Page 19: End of software Dwarf section ==> Good place to break
| |
| − | * Page 20: beginning of hardware
| |
| − | ** processor
| |
| − | ** memory
| |
| − | ** network
| |
| − | * Page 32: Programming models
| |
| − | * Need a new type: '''Unicode''' (words and unicode are important units/types to support!)
| |
| − | ** Does not include threading
| |
| − | ** MPI
| |
| − | ** Map-REduce
| |
| − | * Page 36: system software
| |
| − | ** '''Autotuner''': compilers are not good enough. Parallel programming is so hard, that we should have machines try out all possible optimizations and figure out how to make it work best.
| |
| − |
| |
| − | * Page 40: virtual machines to the rescue
| |
| − | * Page 44: conclusion
| |
| − |
| |
| − |
| |
| | | | |
| | | | |
