scieee Science in your language
[en] (orig)

A framework and methodology for performance prediction of HPC workloads

Abstract

The presented poster outlines an approach for predicting the performance of High-Performance Computing workloads (HPC). By utilizing data gathered from runtime hardware counters across a range of HPC applications and benchmarks, we develop an artificial intelligence model based on ensemble tree algorithms. This model is capable of forecasting the performance of other HPC applications. This work differs from current research by focusing on the granularity of training and prediction. Specifically, our model is developed utilizing individual computation bursts as input samples for training. Through this approach, we prove that a prediction of the instructions per cycle (IPC) metric of unseen applications is possible based on architectural performance counters that can be obtained easily with already used and convenient performance tools.

Read accessible full text

A framework and methodology for performance prediction of HPC workloads

Author: Orteu, Júlia,Clascà Ramírez, Marc,Garcia-Gasulla, Marta,Labarta Mancho, Jesús José,Jennings, Elise
Publisher: Barcelona Supercomputing Center
Year: 2024
Source: https://upcommons.upc.edu/bitstream/2117/428527/1/SODS11-2024-28.pdf
A F amewo k and Me hodology o Pe o mance
P edic ion o HPC Wo kloads
J´
ulia O eu∗, Ma c Clasc`
a∗, Ma a Ga cia-Gasulla∗, Jes´
us Laba a∗†, Elise Jennings‡
∗Ba celona Supe compu ing Cen e , Ba celona, Spain
†Uni e si a Poli `
ecnica de Ca alunya, Ba celona, Spain
‡Pa Tec AG, M¨
unchen, Ge many
E-mail: {julia.o eu, ma c.clasca, ma a.ga cia, jesus.laba a}@bsc.es
elise.jennings@pa - ec.com
Keywo ds—HPC Wo kloads, Pe o mance P edic ion, Run ime
Ha dwa e Coun e s, Ins uc ions pe Cycle (IPC), Pe o mance
Tools, Pa allel Applica ions, Reg ession ees, ML & AI
I. EXTENDED ABSTRACT
A. In oduc ion
The p esen ed pos e ou lines an app oach o p edic ing
he pe o mance o High-Pe o mance Compu ing wo kloads
(HPC). By u ilizing da a ga he ed om un ime ha dwa e
coun e s ac oss a ange o HPC applica ions and benchma ks,
we de elop an a i icial in elligence model based on ensemble
ee algo i hms. This model is capable o o ecas ing he
pe o mance o o he HPC applica ions. This wo k di e s om
cu en esea ch by ocusing on he g anula i y o aining
and p edic ion. Speci ically, ou model is de eloped u ilizing
indi idual compu a ion bu s s as inpu samples o aining.
Th ough his app oach, we p o e ha a p edic ion o he
ins uc ions pe cycle (IPC) me ic o unseen applica ions
is possible based on a chi ec u al pe o mance coun e s ha
can be ob ained easily wi h al eady used and con enien
pe o mance ools.
B. Resea ch and De elopmen
The esea ch line ocuses on explo ing he possibili y o
p edic ing he pe o mance, and ul ima ely he execu ion ime,
o known wo kloads in HPC machines p io o hei execu ion.
P edic ing he pe o mance o HPC applica ions o
wo kloads is a complex ask wi h a widely discussed se
o app oaches and me hodologies. The la es esea ch wo k
on his opic ocuses on da a-d i en me hodologies using
machine lea ning, bu p e ious s udies p esen di e en ypes
o p edic ion me hods: analy ical me hods, ha can be de i ed
om a machine ep esen a ion, o can be he esul o a
s a is ical wo k; and non-analy ical me hods, which comp ise
he a i icial in elligence echniques and he simula ion
me hods [1] .
Ou app oach dis inguishes i sel om exis ing esea ch
by i s ocus on he g anula i y o aining and p edic ion.
Using pe o mance analysis ools Ex ae and Pa a e
[2] [3], de eloped in Ba celona Supe compu ing Cen e
(BSC), we ex ac da a a he compu a ional bu s le el and
a ge he IPC me ic o e e y bu s as a pe o mance measu e.
Fig. 1. Ex ae and Pa a e gene aliza ion o bu s
A bu s is de ined as he ime in e al o ac i e ope a ion
be ween wo successi e e en s in a p ocess. In Figu e 1,
we p o ide a isual example o a imeline om a execu ion
ep esen a ion o an applica ion unning wo p ocesses. The
ac i i ies wi hin hese p ocesses ha e been ca ego ized in o
wo ypes: communica ion bu s s, indica ed in ed, and com-
pu a ion bu s s, ep esen ed in blue. These ca ego iza ions a e
de i ed om MPI e en ma ke s, wi h g een lags ini ia ing
compu a ion and ed lags ma king he s a o communica ion.
In ou s udy, we concen a e on compu a ion bu s s, which
we e e o as use ul bu s s. These bu s s a e pe iods whe e
he applica ion is ac i ely engaged in da a p ocessing and
execu ing ins uc ions. These indi idual use ul bu s s se e
as he inpu samples o he model, as one en y o ou
da ase , highligh ing ha his app oach p o ides a e y p ecise
pe o mance p edic ion o a e y speci ic applica ion’s pa .
C. Fea u es and wo kload cha ac e iza ion
To acili a e ou s udy, we conside a se o Pe o mance
Applica ion P og amming In e ace (PAPI) coun e s [4] as he
ounda ional da a. These coun e s include he o al numbe o
ins uc ions comple ed (N), he o al cycles seen by h ead
(Cyc), memo y load ins uc ions (NLD), memo y s o e in-
s uc ions (NSR), b anch ins uc ions (NBR), and o al cache
miss e en s a bo h L1 and L3 le els (missL1,missL3).
We no malize he da a in each bu s by using he a io o
ins uc ions and cache misses ins ead o absolu e coun e s
alues. This allows us o compa e any indi idual bu s om
any pa o a ace and om any applica ion. F om he
coun e s, we cha ac e ize he a iables using he ollowing
compu a ions:
•Ins uc ion mixes, which a e a ios o speci ic ins uc-
ion ypes o he o al numbe o ins uc ions, such as
LD =NLD
N o loads, SR =NSR
N o s o es, and
BR =NBR
N o b anches.
•Cache miss a es, calcula ed as L1=missL1
NLD+NSR o
he L1 cache and L3=missL3
NLD+NSR o he L3 cache,
which a e indica o s o memo y access e iciency.
•The a e age node concu ency du ing co e execu ion,
h ough he in eg al o he Pa allelism unc ion o e
he ime om Tbegin o Tend:RTend
Tbegin P a ( )d .
•IPC, which is he a ge pe o mance me ic, gi en by
he equa ion IP C =N
Cyc .
D. Da a sou ces
The p ocess in ol es he selec ion o a speci ic se o
benchma ks and ke nels o ex ac da a and adap i o he
pu pose o aining he models. Addi ionally, a sepa a e se o
applica ions has been chosen o e alua ing he pe o mance
o he ained models ( es ing).
The selec ion o ke nels and benchma ks o he aining
se is a pi o al decision ha acili a e he ep esen a ion o he
bu s space, enabling he gene aliza ion o new applica ions.
Fo each applica ion selec ed o he aining se , we’ e a ied
he p oblem sizes o cap u e a comp ehensi e da ase . The
na u e o he a ia ion depends on he applica ion’s cha ac e is-
ics—i could be he size o an a ay, he g anula i y o a mesh,
o he complexi y o inpu s. Addi ionally, we’ e scaled he
compu a ional wo kload by al e ing he numbe o p ocesses
wi hin a single node o each a ian o p oblem size. This
me hodical app oach allows us o cons uc a aining da ase
ha co e s a wide ange o scena ios.
E. Da a ex ac ion F amewo k
Fig. 2. Flowcha o he da a ex ac ion pa o he aining amewo k
The da a ex ac ion p ocess o each applica ion, bo h o
aining and es ing da ase s, is ou lined in Figu e 2.
The depic ed da a ex ac ion p ocess begins wi h he execu-
ion o a known applica ion on a known machine. In his case,
we ha e used he Ma eNos um 4 supe compu e ’s a chi ec u e
o e e ence. We ob ain he use ul bu s s om a p og am
execu ion using he BSC ools, which a e hen p ocessed in o
a ea u es o ma .
F. AI Model
The s udy in es iga es a ange o di e se machine lea ning
algo i hms wi h he aim o aining an e ec i e p edic i e
model. In his explo a ion, we ha e employed a 10- old
c oss- alida ion me hod on ou aining da ase o e alua e
he pe o mance and compa ibili y o hese algo i hms
wi h ou ype o da a. The empi ical esul s highligh a
clea ad an age o using Boos ing Ensembles ha ely
on decision ees, leading o a ocus on XGBoos me hod
[5] . This also ma ches wi h p e ious esea ch conclusions [6].
To ensu e a ai assessmen o he models, we’ e de eloped
a me hod o injec hese p edic ions in o Pa a e aces and
we’ e de ised speci ic e o me ics in o he ace ailo ed o
e alua e he pe o mance ou comes o each model. This allows
o a mo e p ecise analysis o how well he models p edic
applica ion pe o mance.
G. Summa y
The pos e shows he main indings and discusses ou ex-
plo a ion o his opic. We p esen a me hod o da a collec ion
and p ep ocessing based on low-e o p og am ins umen a-
ion and au oma ic ools, as i is well known in he li e a u e
ha being able o collec da a au oma ically and building he
model e o lessly is c i ical o end up wi h aluable and
con enien aining and p edic ion wo k low.
We also s udied how changing he ea u e se , he aining
da a size o he machine lea ning algo i hm a ec s he accu-
acy. We discuss a me hod o cha ac e ize compu a ional bu s s
based on ins uc ion mix ea u es and ins an aneous machine
concu ency ha is la e able o classi y, using ees, he IPC
o unseen bu s s. The e o e, we p o e ha i is possible o
o esee he pe o mance o a whole unseen applica ion ace
based on his cha ac e iza ion me hod.
II. ACKNOWLEDGMENT
This wo k bas been published in p oceedings o he 11 h
In e na ional BSC Se e o Ochoa Doc o al Symposium, 2024.
REFERENCES
[1] J. Flo es-Con e as e al., “Pe o mance p edic ion o pa allel applica-
ions: a sys ema ic li e a u e e iew,” The Jou nal o Supe compu ing,
ol. 77, no. 4, pp. 4014–4055, 2021.
[2] V. Pille e al., “Pa a e : A ool o isualize and analyze pa allel code,” in
P oceedings o WoTUG-18: anspu e and occam de elopmen s, ol. 44,
1995, pp. 17–31.
[3] H. Se a e al., “F amewo k o a p oduc i e pe o mance op imiza ion,”
Pa allel Compu ing, ol. 39, no. 8, pp. 336–353, 2013.
[4] S. B owne e al., “Papi: A po able in e ace o ha dwa e pe o mance
coun e s,” 1999.
[5] T. Chen and C. Gues in, “Xgboos : A scalable ee boos ing sys em,”
in P oceedings o he 22nd ACM SIGKDD In e na ional Con e ence on
Knowledge Disco e y and Da a Mining, se . KDD ’16. New Yo k, NY,
USA: Associa ion o Compu ing Machine y, 2016, p. 785–794.
[6] J. Sun e al., “Au oma ed pe o mance modeling o hpc applica ions using
machine lea ning,” IEEE T ansac ions on Compu e s, ol. 69, no. 5, pp.
749–763, 2020.
J´
ulia O eu is a s uden in he i s coho o he
Bachelo ’s deg ee in A i icial In elligence a Uni-
e si a Poli `
ecnica de Ca alunya (UPC). Since 2023,
has been a Junio Resea ch Enginee a he Ba celona
Supe compu ing Cen e (BSC) in he Bes P ac-
ices o Pe o mance and P og ammabili y (BePPP)
g oup, whe e she cu en ly wo ks as a pe o mance
p edic ion analys o HPC Applica ion Wo kloads.