You are on page 1of 4

Erasmus Mundus Master

Distributed Computing
Spring 2014
Parallel Programming Tools and Models
Assignment 3 Analysing P!!T appli"ation
1 Introduction
#e are going to analy$e t%e tra"es o& t%e P!!T appli"ation' P!!T is an
implementation o& a mesos"ale mi"%rome"%ani"al material model' At t%e %eart o& t%e
appli"ation lies a set o& !ast !ourier Trans&ormations t%at signi&i"antly redu"e t%e
solution time' (y en"ompassing a large number o& numeri"al met%ods )it% *arying
memory and CP+ utili$ation patterns P!!T is representati*e to aspe"ts o& algorit%ms
used in s"ienti&i" "omputing in general'
2 Measurements
MPI_Stats
#e obtain tra"es &rom appli"ation runs )it% 2, 4, -, 1., 32, .4 and 12- pro"essors'
A&ter)ards t%e tra"e &iles be"ome bigger t%an 4gb /despite e&&orts to limit it0 t%ere&ore
*ery di&&i"ult to pro"ess' T%e problem si$e is a grid o& 21.221.221.' Ea"% pro"essor is
assigned one point in t%e 3 a2is'
# of procsRuntime (us)
2 7482403683
4 3804607196
8 1944545338
16 978807873
32 521907615
64 280413148
128 193743188
4 8 16 32 64 128
0
5
10
15
20
25
30
35
40
45
Speed Up
Speed Up
#e noti"e t%at t%e appli"ation s"ales in an almost linear &asion until 32 pro"essors' A&ter
t%is point s"alability ta4es a %it )%i"% is e*en more noti"eable in t%e "ase o& 12-
pro"essors'
#e "ontinue our analysis by e2amining t%e e&&i"ien"y model o& t%e appli"ation' #e
e2tra"t t%is in&ormation &rom t%e MP56stats "on&iguration &ile and t%e Dimemas
simulation'
#e noti"e t%at t%e results o& t%e e&&i"ien"y model mat"% )%at )e %a*e already seen &rom
t%e runtime measurements' 7e*ert%eless t%e %it in parallel e&&i"ien"y )%en )e 8ump
&rom .4 to 12- "ores is 9uite pro&ound'
2 4 8 16 32 64 128
0
1000000000
2000000000
3000000000
4000000000
5000000000
6000000000
7000000000
8000000000
Runtime
Runtime
# of procsParallel Efficienc !omm Efficienc "oa# $alance %"$
2 98&49 98&78 100 98&78
4 97&63 98&03 100 98&47
8 96&15 97&56 99 98&14
16 95&28 96&69 99 97&64
32 92&91 94&09 99 96&14
64 87&81 89&85 98 95&36
128 63&9 67&33 95 72&45
2 Application Structure
#e "ontinue by analy$ing t%e stru"ture o& t%e appli"ation' #e are loo4ing &or t%e
e2pe"ted iterations as e2pe"ted by t%e sour"e "ode o& t%e appli"ation'
2 4 8 16 32 64 128
0
20
40
60
80
100
120
Efficiency
Parallel Efficiency
Comm Efficiency
oad !alance
"!
#ran$fer
T%e same pattern "ontinues t%roug%out t%e di&&erent si$es'
#e "ontinue by loading t%e use&ull duration "on&iguration'
#e noti"e t%at a big part o& t%e "omputation is spent )it% lo) use&ul duration *alues'
T%is is a problem t%at re9uires some &urt%er e2amination'
Useful duration: 8 cores

You might also like