Last Friday, Dr Ioana Banicescu of computer science dept. of Mississippi State University gave us a very nice seminar on the experiences with fractiling in N-body simulations.

At first, Dr Banicescu introduced the overview: N-body problem; FMA; critical performance factors, scheduling via fractility; implementation on IBM sp2 & super MS; etc.

Then Dr Banicescu presented N-body problem: O(N^2) naive algorithm, approximation algorithms; the green guard algorithm; traveling the tree upwards.

Then Dr Banicescu introduced the parallelization: distribute subtrees among processors, performance issues; mapping; computation time and accuracy; load balance; communication; related works.

Then Dr Banicescu presented fractiling (algorithm, example); tiling; implementation; architecture independent considerations; the KSR1 architecture (advantage, disadvantage, hardware); the IBM-SP2 architecture; implementation.

Then Dr Banicescu introduced the master/slaves communication; implementation highlights; state duration for 4 processors; Gaussian distribution; Corner distribution; 3-D experimental results: fract/PFMA improvement % KSR1, PFMA/fract C.O.V. (KSR1).

Then Dr Banicescu presented PFMA/fractiling variance; 3-D cost: uniform (KSR1/corner/Gaussian); 3-D conclusions (KSR1).

At last Dr Banicescu introduced the experimental results: (fract/PFMA), improvement % (super MSPARC), average C.O.V. (IBM-SP2/super-MSPARC), cost (uniform/Gaussian/corner PxTp), performance gains summary, and conclusions.

Last Update: 5/14/98
Web Author: Zizhong Wang
The report is for Dr Paprzycki@ marcin.paprzycki@ibspan.waw.pl or@ Home Page