Report on Vectorization Application Review

S.C. Jardin

Princeton Plasma Physics Laboratory
Princeton University

Table of contents

  1. Present NERSC Configuration
  2. Results of PPPL Codes Survey
  3. Possible Solutions
  4. Summary


Present NERSC Configuration

Number of ProcessorsMemory per processor (MW)Time Relative to C-90
T3E 120 32(D) 1-11+
C-90 16 16(S) 1
J90-K 32 32(S) 3-4
J90-F 32 32(S) 3-4
J90-B 32 32(S) 3-4
T932* 32 32(S) .5
* Not presently planned for NERSC
+ No penalty assumed for distributive memory

top



Results of PPPL Codes Survey

Present Run
mem(MW) time(hr) 		Parameters Needed
mem(MW) time(hr) 		Runtime Factor
DEC-alpha / C-90
small codes0 1 -- 0.1 1 -- 0.5 2
2D MHD1 20 -- 5 80 -- 40 6
3D MHD2 40 -- 16 320 -- 350 11
3D PIC3 40 -- 40 400 -- 400 11
Monte-Carlo4 1 -- 1 1 -- 60 1
Linear Eigenvalue5 40 -- 0.1 400 -- 1 1

E.g.:

0FULL
1TSC
2MH3D, GYROFLUID
3ORBIT, MH3D-K, Gyrokinetic
4DEGAS2
5PEST, SPARK
(All of these categories are of comparable importance)

top



Possible Solutions

2D + time MHD Codes: 80MW x 40 Hours

platform #P computation hours fraction of machine hours x fraction
C-90 5 (40x1/5) 8 0.31 2.48
J-90 3 (40x3/3) 40 0.10 4.00
T3E 3 (40x6/3) 80* 0.025 2.00*
T932 3 (40x.5/3) 6 0.10 0.60
*assumes linear speedup

3D + time MHD Codes: 320 MW x 350 Hours

platform #P computation hours fraction of machine hours x fraction
C-90 201 (350)x1/20) 17 1.25 21
J-90 10 (350x3/10) 100 0.31 31
T3E 10 (350)x11/10) 385* 0.083 32*
T932 10 (350x.5/10) 17 0.31 5.3
*assumes linear speedup
1not possible

3D + time PIC CODES: 400 MW x 400 Hours

platform #P computation hours fraction of machine hours x fraction
C-90 251 (400x1/25) 16 1.56 25
J-90 13 (400x3/13) 92 0.40 37
T3E 13 (400x11/13) 340* 0.108 37*
T932 13 (400x.5/13) 15 0.40 6
*assumes linear speedup
1not possible

3D Monte-Carlo: 1 MW x 60 Hours

platform #P computation hours fraction of machine hours x fraction
C-90 1 (60x1) 60 .0625 3.75
J-90 1 (60x3) 180 .03125 5.62
T3E 1 (60x1) 60 .0083 0.50
T932 1 (60x.5) 30 .0312 0.94

top



Summary

Equivalent whole machine hours/run for desired parameters

C90 J90 T3E T932
2D + time MHD 2.5 4 2 0.6
3D + time MHD - 30 30 5
3D + time PIC - 40 40 6
Monte Carlo 4 6 0.5 1

top