It’s all MOOT An MPI adventure. Brian F. Haag CSE 633 Dr. Miller The Problem •  CalculaDng basic geneDc algorithms on a cluster. •  Variable dataset, size of a populaDon varies over Dme. •  UlDmately though, the data is very small, but needs to be updated very frequently. Challenges •  Learning C, I’m most comfortable with LISP, but that didn’t pan out, so I switched early. •  Keeping within scope, runDmes for GAs can spiral out of control quickly. •  Improving performance from iniDal versions. What is a Mooter? My parent’s didn’t explain it to me like this. Fitness and Goals •  The fitness goal is to have a populaDon where 80%+ of a populaDon greater than the iniDal populaDon survive fitness tests for 5 consecuDve fitness tests. •  A fitness test is a test against the stats of a given Mooter, if the Mooter ‘s stats are not up to snuff, then they do not survive and their genes do not conDnue forward. •  Every x generaDons, Mooters are not limited to their township in terms of potenDal maDng partners(Great Moots). Example Technique •  The populaDon is sca_ered through the cluster, with each cluster taking a porDon relaDve to the size of the populaDon •  Wanderlust changes are broadcast out •  IniDal populaDon is staDc to making tesDng easier. ObservaDons •  Increasing starDng wanderlust has a dramaDc (negaDve)impact on performance. •  Modeling parallel GA using “realisDc” scenarios is actually pre_y damn hard, there are a lot of potenDal sequenDal steps once you start dividing the populaDon(well giving them some mobility). •  Cont ObservaDons Cont •  Increasing the iniDal fitness thresholds can result in dramaDcally speeding up the goal. •  Increasing the rate of Great Moots will increase performance •  ImplemenDng a parallel sorDng mechanism would likely improve ability to scale program IniDal TesDng 1 core per nodes 12 Moot Rate: 100 generaDons PopulaDon Size: 100k Running Dme in minutes. Running Time 10 8 6 4 2 PEs 0 0 10 Running Dme average of 5 tests. 20 30 PEs 40 50 PE 10.42 7.38 5.12 3.22 2.46 2.54 2.41 60 70 1 2 4 8 16 32 64