www.bsc.es

Jesús LabartaBSC

Performance Analytics From visualization to insight

VI-HPC 10th Anniversary WorkshopFrankfurt, June 23rd 2017

BSC Tools framework

CUBE, gnuplot, vi…

.prv+

.pcf

.dim

Time Analysis, filters

.cfg

Paramedir

.prv

Instruction level simulators

Tasksim, Sniper, gem5

XMLcontrol

ParaverLD_PREPLOAD,

Valgrind, Dyninst, PINPAPI

MRNETExtrae

DIMEMAS

VENUS (IBM-ZRL)

Machine description

Trace handling & displaySimulatorsAnalytics

Open Source http://www.bsc.es/paraver

The importance of detail and flexibilitySince ~ 1991

Performance analytics

prv2dim

.xls.txt

.cube

.plot

Instr. traces

3

Performance Analytics

Performance analysis: THE big data app– 10000 cores x 1 event/100us x 100 bytes/event x 1000s = 10 GB– Easily underestimated terms by orders of magnitude

Performance Analytics:– Squeezing the information in the captured data Insight

• About Machine Learning … and beyond, and before– Some BSC activities in the last 10 year (~)

Vision– Have to leverage data processing methods from ALL areas– Have to balance between first principles, pure statistical, black box

4

BSC Performance Analytics

Instantaneous metrics for ALL hardware counters at “no” costAdaptive burst mode tracing

Tracking performance evolution

26.7MB traceEff: 0.43; LB: 0.52; Comm:0.81

1600

cor

es

2.5 s

BS

C-E

S –

EC

-EA

RTH

BS

C-E

S –

EC

-EA

RTH

AM

G20

13

Flexible trace visualization and analysis

Advanced clustering algorithms

5

BSC Performance Analytics

eff.csv

Several core counts

eff_factors.py

extrapolation.py

Dimemas

No MPI noise + No OS noise

“Scalability prediction for fundamental performance factors ” J. Labarta et al. SuperFRI 2014

Models and Projection Data access patterns

Runtime Energy and performance model (EAR)

Inte

l –B

SC

Exa

scal

eLa

b

Leno

vo -

BS

C

G. Llort et al, “Scalable tracing with dynamic levels of detail” ICPADS 2011

Structure detection

About Analytics AND infrastructure

T0

ClusteringAnalysis

MRNetFront-end

T1 Tn

…

Back-end threads

Aggregatedata

Broadcastresults

Structure detection

Intel –BSC Exascale Lab

MRNET

PyCOMPSs

7

What we use: POP

8

Performance Analytics

Leverage methods from ALL areas

Balance between first principles, pure statistical, black box

– “Proper” choice of feature vector

Leverage big data infrastructure

Lots of opportunities for the next 10 years !!!