Towards an Optimisation-Aware Data Centre Prediction Toolkit

Towards an Optimisation-Aware Data Centre Prediction Toolkit

Management of cloud applications and data centre resources has become increasingly complex, due in large part to a substantial increase in the degree of heterogeneity and scale. Topology optimisation tools and algorithms currently in use for optimal placement of applications across such heterogeneous resources are typically trial-and-error based. Predicions enable the evaluation of such optimisation algorithms and their complex interactions towards better reasoning at decision time, e.g. optimising for a specific trade-off.

This  project  has  received  funding  from  the  European  Union’s  Seventh  Framework  Programme  for  research,  technological  development  and  demonstra?on  under  

grant  agreement  no.  610711    

Sim

CactoScale acquires and analyses application behaviour and infrastructure performance data. It consolidates multiple sources of performance and error monitoring data and updates Infrastructure Models accordingly.

CactoOpt applies mathematical models to optimise application-resource mappings and create Optimisation Plans. It includes a library of configurable multi-objective optimisation algorithms to maximise data centre efficiency.

CACTOS Runtime Toolkit

Logical Load Model

Physical Load Model

Logical Data Centre Model

Physical Data Centre

Model

identifiesPhysical

Configuration

measures Load

identifies Logical

Configuration

creates

updates

extractsPhysical State

feeds

CactoOpt

adapts

reconfigures & deploys

CactoScale

CACTOS Prediction Toolkit

VMI OpenStack

VMI FCO

Virt

ualis

atio

n M

iddl

ewar

e In

tegr

atio

n

reconfigures & deploys

OptimisationPlan Model

creates

feeds

updatesCactoSim

Infrastructure Models

optimises

measures Load

identifies Physical and Logical Configuration

simulates

captures

provides Optimisation

Actions

VMI CactoSim

simulatesSystem Load

Real-WorldSimulation

provides Measurements

Data Centre

www.cactosfp7.eu  

2014 Symposium on Software Performance, November 26-28, Stuttgart, Germany

Sergej Svorobej, James Byrne, PJ Byrne Dublin City University, Dublin 9, Ireland Henning Groenda, Christian Stier FZI Forschungszentrum Informatik, Karlsruhe, Germany

IVMIService

OpenStackAPI

CDO Model Generator

Runtime Model Storage

Prediction Model Storage

Cyclic Optimiser

CactoSim EngineCactoSim

IDE (Eclipse-based)

HBase

SQL DB

SQL DB

Chukwa

CDO

OptimizationEngine

CACTOS Chukwa Agent

CSE API

EMF Store

Virtualisation Middleware Integration

Control

FCOAPI

Conforms Provides Component Type (CPCT)

CPCT

CSEConnector

CACTOS Runtime Toolkit Version 1

COSControl

controlChukwaCollector

HBaseZooKeeperHadoop

Pig

CactoScale

CactoOpt Infrastructure

Optimiser

CactoOptIntegration

CACTOS Prediction Toolkit Version 1

CactoSim

OfflineAnalysis

VMIFCO

VMIOpenStack

CactoOptKnowledge DB

VMICactoSim

CyclicOptimiserSimulation

CactoSim’s Architecture is driven by the integration of the CACTOS Prediction Toolkit and the CACTOS Runtime Toolkit. CactoSim Engine and IDE build on Palladio and the SimuLizar Plugin for analyzing self-adaptive systems. They use a more complex Infrastructure Model and employ QVTo model transformations internally to map between the PCM and this model. The Prediction Model Storage uses EMF Store for versioning different data centre design and configuration alternatives. The Cyclic Optimiser Simulation uses this mapping to delegate optimisation plan creation to the CactoOpt Infrastructure Optimiser. This plan is enacted by VMI CactoSim within a simulation run.

Colors:  CactoScale:  175  40  30  CactoOpt  60  150  240  CactoSim  190  190  40  Cactos  141  198  63  Green  Lines  190  190  40  

CactoSim is a discrete event simulation (DES) framework that enables the evaluation of the effect of provisioned CactoOpt runtime optimization strategies in a simulated environment at design time. This offers significant benefit over the utilization of testbeds which are typically costly to provide on a large scale and can have a high level of complexity. CactoSim allows for reproducible and controlled experimentation with different workload mix and resource scenarios enabling both performance and risk analysis to be performed as well as the tuning of potential bottlenecks in a data centre.

The current prototype integrates with the Runtime Toolkit and enables detailed data centre modelling. The simulation provides results for decision support through seamless model transformation to PCM. Further iterations will improve the mapping and prediction factors.