Some Lessons from Applications for High Performance Fortran(HPF) and its Extensions -- High Performance Fortran II Kickoff Meeting

January 13,1994 Houston Texas Geoffrey Fox This talk was given at the CRPC sponsored HPF II Kickoff meeting hosted by Rice University. Forms of Parallelism are classified and some 21 application areas examined in more or less detail for the posssible use of HPF or hpf= -- this refers informally to set of proposed HPF extensions.
The slides (Click on number for slide, or text for descriptions):

.1=..Title Slide for 94b: Some lessons from applications for high performance FORTRAN (HPF) and its extensions: High performance FORTRAN II kickoff meeting.
.2=..Goal of HPF and lessons used in its design.
.3=..What applications does HPF and HPF+ support? Expressed in terms of problem architecture classification of Fox.
.4=..The five major problem architectures and their qualitative parallelism and HPF nature.
.5=..The three type of parallelism that any complete programming environment must handle.
.6=..What issues must or could be addressed in HPF+: Extensions to HPF.
.7=..HPF an HPF+ for a set of applications - mainly from partial differential equation and particle dynamics areas.
.8=..Detailed description of HPF successes and failures: Current HPF can do: A set of synchronous problems.
.9=..Detailed description of HPF successes and failures: HPF can also do the synchronous order N squared particle dynamics.
10=..Detailed description of HPF successes and failures: Current HPF can also do - Embarrassingly parallel problems.
11=..Detailed description of HPF successes and failures: difficult but (almost) possible for HPF.
12=..Detailed description of HPF successes and failures: HPF can express region growing in image processing.
13=..Details successes and failures: HPF can also express irregular domains seen near critical points of physical systems.
14=..Example of clustering near a critical point: Swedsen - Wang clusters (boundaries shown in black) for 3 state Potts model at Tc
15=..Detailed description of HPF successes and failures: Significant improvement in HPF needed but seems possible for irregular (adaptive) loosely synchronous problems such as PIC and ICRM.
16=..Particle in the cell (PIC) and irregular coupled regular meshes (ICRM).
17=..Detailed successes and failures: Some very hard loosely synchronous problems such as direct matrix solvers and fast math pole methods.
18=..10,000 body Barnes-Hut Tree.
19=..8 million bodies - 10 Mpc diameter final state with ~800 resolved "galaxies" (Warren, Quinn, Zerek).
20=..The largest "galaxy" halo (137,000 bodies) from the 8.8 million simulation.
21=..Preliminary summary: HPF features for different application fragments I: Regular gnd to FFT.
22=..Preliminary summary: HPF features for different application fragments II: Image processing to domain decomposition.
23=..Preliminary summary: Particle in the cell to direct sparse solvers.

InfoWarehouse Management team