[PPT] - Analytical models for performance and energy consumption evaluation PowerPoint Presentation

SLIDE 1

Analytical models for performance and energy consumption evaluation of storage devices

Eric Borba

Universidade Federal de Pernambuco (UFPE) erb@cin.ufpe.br

October, 2020

SLIDE 2

Agenda

Introduction Methodology Models Conclusion and future works

EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

Experimental results

SLIDE 3

Introduction Methodology Models Experimental Results Conclusion Context Motivation Goal Why GSPN?

Context

Cloud computing (explosive growth of data – Big data)
90% of all data in the world (2018-2020) -> 163ZB (2025)
Storage systems represent:

25-35% of the Cloud computing costs 13% energy consumption in data centers 90% of a transaction execution time

EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 4

Context

HDDs causes higher latency due to mechanical

positioning in random access

Low capacity, shorter lifetime, and cost of SSDs are

some obstacles

Alternatively, hybrid (SSD+HDD) approaches have

been proposed

Introduction Methodology Models Experimental Results Conclusion Context Motivation Goal Why GSPN? EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 5

Motivation

Aspects, such as performance and energy consumption,

must be balanced

Few works rely on GSPN (generalized stochastic Petri nets)

and comprise both metrics

Effective arrangement through optimized data-placement

policies

Introduction Methodology Models Experimental Results Conclusion Context Motivation Goal Why GSPN? EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 6

Goal

”Conceive stochastic models to estimate energy consumption and performance of storage systems”

Introduction Methodology Models Experimental Results Conclusion Context Motivation Goal Why GSPN? EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 7

Why GSPN?

Introduction Methodology Models Experimental Results Conclusion Context Motivation Goal Why GSPN?

Synchronization, resource sharing, conflicts
Non-exponential activities and zero delays (logical control)
Alternative to Markov chains generation (simulation

techniques)

Analysis of quantitative and qualitative properties

EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 8

Introduction Methodology Models Experimental Results Conclusion Methodology Design of Experiments Tools and environment setting

Methodology

Analytical models
Measurement step: moment matching/validation
Factorial design (20 replications): screening experiment
Three experiments: SPC (storage performance council)
Metrics: IOPS (input/output per second)/energy

consumption, and price/IOPS

EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 9

Introduction Methodology Models Experimental Results Conclusion Methodology Design of Experiments Tools and environment setting

Design of experiments

Experiment I Screening approach Main factors and second-order interactions Experiment II Random access (4KB; 70%_write) Experiment III Sequential access (1MB; 50%_write) Experiment IV Mixed access (1MB; 50%_write; 80%_rnd)

EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 10

Tools Iometer (benchmark) Oscilloscopes Environment

Tools and environment setting (measurement step)

Introduction Methodology Models Experimental Results Conclusion Methodology Design of Experiments Tools and environment setting EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 11

Introduction Methodology Models Experimental Results Conclusion Single storage model Multiple storage model

Single storage model

EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 12

Introduction Methodology Models Experimental Results Conclusion Single storage model Multiple storage model

Multiple storage model

EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 13

Introduction Methodology Models Experimental Results Conclusion MomentMatching/Validation Experiment I: screening Experiment II: random access Experiment III: sequential access Experiment IV: mixed access

Moment matching – phase type

EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 14

Introduction Methodology Models Experimental Results Conclusion MomentMatching/Validation Experiment I: screening Experiment II: random access Experiment III: sequential access Experiment IV: mixed access

Validation

EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 15

Introduction Methodology Models Experimental Results Conclusion MomentMatching/Validation Experiment I: screening Experiment II: random access Experiment III: sequential access Experiment IV: mixed access

Experiment I: screening

EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 16

Introduction Methodology Models Experimental Results Conclusion MomentMatching/Validation Experiment I: screening Experiment II: random access Experiment III: sequential access Experiment IV: mixed access

Experiment II: random access

EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 17

Experiment III: sequential access

Introduction Methodology Models Experimental Results Conclusion MomentMatching/Validation Experiment I: screening Experiment II: random access Experiment III: sequential access Experiment IV: mixed access EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 18

Experiment IV: mixed access

Introduction Methodology Models Experimental Results Conclusion MomentMatching/Validation Experiment I: screening Experiment II: random access Experiment III: sequential access Experiment IV: mixed access EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 19

Remarks

HDDs performance issues: small objects, random

access, simultaneous requests

For sequential accesses, and large objects, HDDs

are still a feasible option

SSDs: suitable for small random readings
Hybrid: for systems in which performance

requirements prevail over energy consumption

Introduction Methodology Models Experimental Results Conclusion Remarks Conclusion Future works EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 20

Conclusion

An approach based on GSPN for assessment of

storages

Evaluation of different technologies and workloads
Experiments illustrate the practical feasibility

Introduction Methodology Models Experimental Results Conclusion Remarks Conclusion Future works EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 21

Future Works

”As future work, we are developing models for assessing the reliability and availability of storage systems”

Introduction Methodology Models Experimental Results Conclusion Remarks Conclusion Future works EricBorba (erb@cin.ufpe.br) Analytical models for performance and energy consumption evaluation of storage devices

SLIDE 22