Difference between revisions of "CSC352 Notes on A View From Berkeley"
| (6 intermediate revisions by the same user not shown) | |||
| Line 1: | Line 1: | ||
<onlydft> | <onlydft> | ||
| − | + | ==Landscape of parallel computing research: a view from berkeley== | |
---- | ---- | ||
* good intro paper. | * good intro paper. | ||
| Line 33: | Line 33: | ||
* '''The word is a new unit of information''' | * '''The word is a new unit of information''' | ||
* Page 11: 3 new areas for adoption as dwarf? | * Page 11: 3 new areas for adoption as dwarf? | ||
| − | ** machine learning | + | ** machine learning (by the way, N-Queens is an example of backtracking algorithms) |
| − | ** database software | + | ** database software (Map-Reduce). Large databases are important for future. |
*** Page 12: map-reduce | *** Page 12: map-reduce | ||
** computer graphics | ** computer graphics | ||
| Line 43: | Line 43: | ||
* Page 20: beginning of hardware | * Page 20: beginning of hardware | ||
** processor | ** processor | ||
| − | ** memory | + | *** Power is the important limiting factor right now |
| + | *** Physics becomes important | ||
| + | *** upper limits: peak: 150W for server, 40W for laptop, 2W for embedded applications. | ||
| + | * ==> '''Art versus Science''' | ||
| + | * Page 22: RISC are returning/should return to simple architecture | ||
| + | * heterogeneous processing is important for I/O and OS operations ==> many cores is better than multi core. | ||
| + | ** memory coherency is important. Must adopt solutions for it. | ||
| + | * Page 28: transactional memory: 1993: '''Old solutions may come back as feasible.''' | ||
| + | *** Memory must be redesigned. Gap between memory and processors is widening. | ||
** network | ** network | ||
| + | * Page 31: it's hard to move intelligence away from human beings. | ||
| + | * art/science again | ||
* Page 32: Programming models | * Page 32: Programming models | ||
* Need a new type: '''Unicode''' (words and unicode are important units/types to support!) | * Need a new type: '''Unicode''' (words and unicode are important units/types to support!) | ||
| Line 50: | Line 60: | ||
** MPI | ** MPI | ||
** Map-REduce | ** Map-REduce | ||
| + | * Page 33: ''' "we believe that integrating research on human psychology and problem solving into the broad problem of designing, programming, debugging, and maintaining complex parallel systems will be critical to developing broadly successful parallel programming models and environments". ''' | ||
| + | * Page 34: MPI will fail. | ||
| + | * Page 35: Unicode!!!! important new type. | ||
| + | * Page 35: '''Some styles of parallelism have proven successful for some applications and no style has proven best for all.''' | ||
* Page 36: system software | * 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. | ** '''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. | ||
| + | ** Autotuner = fancy search | ||
| + | |||
* Page 40: virtual machines to the rescue | * Page 40: virtual machines to the rescue | ||
| + | * Page 42: No efficient metrics to measure performance... Hardware counters to the rescue. | ||
* Page 44: conclusion | * Page 44: conclusion | ||
| + | ==PowerPoint from the Web== | ||
| + | |||
| + | * http://www.citris-uc.org/system/files/Kathy-Yelick-ITRI_CITRIS.pdf | ||
| + | |||
| + | ==Student Papers== | ||
| + | * [[Media:CSC352_StudentReviewAViewFromBerkeley1.pdf | paper1.pdf ]] | ||
| + | * [[Media:CSC352_StudentReviewAViewFromBerkeley2.pdf | paper2.pdf ]] | ||
| + | * [[Media:CSC352_StudentReviewAViewFromBerkeley3.pdf | paper3.pdf ]] | ||
| + | |||
| + | =Notes from 2013= | ||
| + | * Lookup Kathy-Yelick video or presentation | ||
| + | * [http://islped.org/X2007/PattersonISLPED07.pdf Kathy Yelick] presentation | ||
| + | * Do not use benchmark, use Dwarfs | ||
| + | * Look at the organization of the paper | ||
| + | * "We concluded that sneaking up on the problem of parallelism via multicore solutions was | ||
| + | likely to fail and we desperately need a new solution for parallel hardware and software. " (page 2) | ||
| + | * ==> Multicore | ||
| + | * Organization: Page 4: The organization of the report follows the seven questions of Figure 1. ... | ||
| + | * look at citation[name year] | ||
| + | * Monte Carlo: compute pi | ||
</onlydft> | </onlydft> | ||
[[Category:CSC352]][[Category:Class Notes]] | [[Category:CSC352]][[Category:Class Notes]] | ||
