Experimental Study of Adaptive Application-Sensitive Partitioning Strategies for SAMR Applications

Sumir Chandra, Johan Steensland, Manish Parashar

The Applied Software Systems LaboratoryRutgers University

(submitted to Super Computing 2001)

Need for Adaptive Partitioning No single partitioning scheme performs the best for

all types of applications and systems Optimal partitioning technique depends on input

parameters and application run-time state Partitioning behavior characterized by the tuple

{partitioner, application, computer system} (PAC) PAC quality characterized by 5-component metric –

communication, load imbalance, data migration, partitioning time, partitioning overhead

Octant approach characterizes application/system state

Adaptive meta-partitioner -> fully dynamic PAC

Dynamic Characterization

Characterizing Partitioner Behavior

3 partitioners Space-filling curve based partitioning (SFC) [from GrACE] Geometric multi-level inverse space-filling curve

partitioning with sequence partitioning (G-MISP+SP) [from Vampire]

p-way binary dissection inverse space-filling curve partitioning (pBD-ISP) [from Vampire]

SFC – good load balance, greater communication and data migration overheads, suited for moderate activity dynamics

G-MISP+SP – favors simple communication and speed over data migration, good load balance, computationally expensive

pBD-ISP – fast, low overheads and communication costs, average load balance, suited for greater communication states with lesser emphasis on load balance

Characterizing Application State RM3D – 3-D compressible turbulence application solving

Richtmyer-Meshkov fingering instability Application trace – 800 coarse level time-steps / 200

snap-shots 128*32*32 base grid, 3 levels, regriding every 4 time-

stepsTime-step

OctantState

PartitionerTime-step

OctantState

Partitioner

0 IV G-MISP+SP 137 VIII G-MISP+SP

5 VII G-MISP+SP 162 II pBD-ISP

25 I pBD-ISP 174 V pBD-ISP

106 VI pBD-ISP 201 III G-MISP+SP

Application State (contd.)

Experimental Results Runs performed on IBM SP2 “Blue Horizon” Measure application execution times for adaptive and

individual runs for different number of processorsPartitioner performance for RM3D application on 64

processors

PartitionerRun-time(seconds)

Max. LoadImbalance

(%)

AMREfficiency

(%)

SFC 484.502 24.878 98.8207

G-MISP+SP 405.062 11.3178 98.7778

pBD-ISP 414.952 35.0317 98.8582

“adaptive” 352.824 8.11825 98.7633

Experimental Results (contd.)

Conclusions Structure of adaptive grid hierarchy is used to

characterize current state and determine partitioning requirements

Adaptive partitioning can improve application performance – for 64 processors, improvement is 27.2% over slowest partitioner

Future work integrate application and system sensitive

mechanisms define policies to drive the partitioner recommender

system