Performance Modeling of High Performance Computing [HPC] S. - - PowerPoint PPT Presentation

Performance Modeling of High Performance Computing [HPC]

• S. Amirhossein Abtahizadeh

March 18th , 2016

Section 1 Failed Attempt

2

Section 1 Ontological Performance Modeling

3

Section 1 Ontology Internet Of Things (IoT) Semantic Web Web of Data

4

Section 1 Web of Data Web 1.0: HTTP Web 2.0: Social Networks Web 3.0: Web of Data

5

Section 1 Web of Data Collections of data Distributed among machines How to present data?

6

Section 1

<?xml version="1.0"?> <!DOCTYPE Ontology [ <!ENTITY xsd "http://www.w3.org/XMLSchema#" > ]> <owlx:Ontology owlx:name="http://www.example.org/wine" xmlns:owlx="http://www.w3.org/2003/05/owl-xml"> </owlx:Ontology>

7

Section 1 Web of Data

Resource Description Framework (RDF) Query another dataset (OWL)

To understand what is “Wine”?

8

Section 1 Semantic Web Interconnected Ontologies Languages: RDF, XML, Turtle Query: SPARQL

9

Section 1 Semantic Web Google:

“polymtl gigl professors” 10

Section 1 Semantic Web What about:

“How many professors at polymtl are working

• n research topic cloud computing now?”

11

Section 1 SPARQL For machines! Not human

12

Section 1

PREFIX ex: <http://example.com/exampleOntology#> SELECT ?capital ?country WHERE { ?x ex:cityname ?capital ; ex:isCapitalOf ?y . ?y ex:countryname ?country ; ex:isInContinent ex:Africa . }

13

Section 1 Internet of Things (IoT) Semantic Web Infrastructure Interconnected machines Raspberry PI*

14

* h.ps://www.raspberrypi.org

Section 1 Research Objective: Present “performance” with Ontology

15

Section 1 Research Objective: Share common understanding of Application Performance

16

Section 1 Research Objective: Tackle the ambiguity of WHAT IS PERFORMANCE?

Nice and clear description in terms of OWL XML fields

17

Section 1

Define quality attributes (What is the availability of the system?)

18

QA 1 QA 2 QA 3 QA 4 QA 5

Section 1 Research Methodology: Develop a Cloud-based app Define scenarios Measure performance with OWL Using logical axioms

19

Section 1 Axioms: Inference Response time < 2ms à P = 90%

20

Section 1 Building Ontology: 100+ axioms Cascaded classification 10+ inferred rules

21

Section 1

22

Section 1

23

Section 1 FAILED! OWL/XML is not efficiently designed Performance is subjective

24

Section 2 Performance and Energy Modeling

• f

High Performance Computing (HPC)

25

Section 2 What is HPC? Parallel processing Advanced applications Massive computations Scientific programs

26

Section 2 What is HPC? Super fast transactions Distributed algorithms Message Passing Interface (MPI)

27

Section 2 In the domain of HPC We deal with CPU-intensive apps

Data might be just an array! Computations might be exponential.

28

Section 2 Aggregate computer powers

Clustering at very large-scale “Computing at the speed of innovation!”*

* IBM (www.ibm.com

29

Section 2 Message Passing Interface (MPI)

30

Section 2 MPI is a Library To write parallel programs It provides collective functions

31

Section 2 MPI is available in many languages: C, C++, Java, Python, R

32

Section 2

• When we have networking libraries,

why bother using MPI ?!

• Optimized for performance
• Fastest network transport found
• Within a computer: MPI will use shared memory (not network!)
• Fast cluster interconnects: MPI will use Infinibands, …
• Enforces guarantees (reliable messages, In-Order)
• Think about the problem, forget about the network

33

Section 2

Research Objective: Given a set of input variables:

Network bandwidth, CPU power, Throughput, Disk speed, Memory, …

What is the optimized configuration for the best performance/energy achieved?

34

Section 2 Def (Performance && Energy): Return Multi-objective Optimization

35

Section 2 Why?

• Efficient resource provisioning (what to choose?)
• Predict the changes in your system (what will happen if.. ?)
• Performance becomes part of the design
• Itemized scenarios (what is important?)
• Avoid surprises with performance when deploying

Enterprise reputation (risk management!)

36

Section 2 How?

37

Section 2

38

Dataset ScienDfic ApplicaDon Master

Node 1 Node 2 Node 3 Node 100

MPI Benchmark Model

Section 2 Architecture Not yet accessible! 100 nodes ORACLE Solaris Cluster ORACLE VirtualBox

39

Section 2 Test Architecture Digital Ocean Cloud Platform 10 nodes ORACLE Solaris Cluster MPI4PY

40

Section 2 Scientific Application Schaffer problem

41

Section 2 Run time Your laptop: life time HPC XT-Cluster: less than a minute

42

Section 2

from platypus.algorithms import NSGAII from platypus.core import Problem, evaluator from platypus.types import Real class Schaffer(Problem): def __init__(self): super(Schaffer, self).__init__(1, 2) self.types[:] = Real(-10, 10) @evaluator def evaluate(self, solution): x = solution.variables[:] solution.objectives[:] = [x[0]**2, (x[0]-2)**2] algorithm = NSGAII(Schaffer()) algorithm.run(10000)

43

Section 2 Methodology Memetic Algorithm Recurrent Neural Network Prediction Model

44

Section 2 Memetic Algorithm Genomes: Set of observed values of Performance & Energy

45

Section 2 Memetic Algorithm Genomes: NumPy array

46

Section 2 Memetic Algorithm Fitness Function: Schaffer optimization

47

Section 2 Memetic Algorithm Cross-Over: Alternating-position Operator

48

Section 2 Recurrent Neural Network Bi-directional data Both past and future

49

Section 2 Prediction Model Compare with benchmarks Linear trend estimation Least Square Error

50

Section 2 Measuring Energy Consumption Power-API Physical devices

51

Section 2 Correlation? Between performance & energy Coefficient: +0.217

52

Section 2

• s

53

Section 2

• Trend Estimation: 83.59%

Per slice of one separate run Noise? I’m working on it ..

54

Conclusion Multi-objective optimization

• f Performance & Energy

In High Performance Computing

55

Conclusion Predict performance and energy SAVE MONEY! Show that this approach is scalable

56

Conclusion Cloud resource selection

57