An Improved Model for Live Migration in Data

An Improved Model for Live Migration in Data Centre Simulators Vincenzo De Maio, Gabor Kecskemeti, Radu Prodan Distributed and Parallel Systems University of Innsbruck

Summary • Introduction – Motivation – Background • Model • Design • Implementation • Evaluation • Conclusion and future work

Introduction • Energy efficiency has become a serious issue for virtual data centres – Around 30 -40 B$ in 2007 (Poess et al. , 2008) • Physical machines (PMs) are often under-utilised – 10 -50% utilisation for most of the time (Barroso et al. , 2007) • Maximising utilisation of PMs to increase energy efficiency of data centres • Workload consolidation

Workload consolidation • Mapping VMs on a reduced subset of PMs

Consolidation: pros and cons • PROs: – Increases utilisation of PMs – The new mapping may allow to shut down unutilised PMs, increasing energy efficiency • CONs: – Drawbacks in performance and energy consumption are possible – Necessity of predicting the benefits of a new VM-PM mapping

Motivation • Evaluating which VM consolidation policy is more efficient – Difficult to perform experiments in real Cloud infrastructures • Cloud Simulation

Cloud simulation • We need – Energy prediction models • CPU, network transfers, I/O • VM Migration – Simulation framework

Our contribution • We propose a model for data centre’s consumption considering – CPU, RAM, network transfers, I/O – VM migration • We implement our model in a cloud simulator • We validate our implementation by comparing it with – Real VM migration measurements – Another state-of-art cloud simulator

Summary • Introduction – Motivation – Background • Model • Design • Evaluation • Implementation • Conclusion and future work

CPU model POWER DRAW vs USAGE, CPU 900, 00 800, 00 POWER [W] 700, 00 Strong change of trend at the same level of load 600, 00 500, 00 400, 00 300, 00 200, 00 100, 00% 10, 00% 20, 00% 30, 00% 40, 00% 50, 00% 60, 00% CPU LOAD AMD Opteron Intel Xeon 70, 00% 80, 00% 90, 00% 100, 00%

Piecewise linear modelling for CPU •

Network and I/O modelling •

Live migration • Iterative process • Modifications to memory pages are constantly transmitted to the target Step 1 VM state moved while still running Step 2 Target updated Step 3 • The continuous memory updates result in more iterations Repeat step 2 Step 4 VM suspended Step 5 VM resumed on the target

Migration energy phases • Normal execution phase – Each actor performs its normal operation (no migration decision taken yet) • Initiation phase – From the issuing of the migration to the effective start of migration • Transfer phase – The state of the VM is transferred over the network • Service activation phase – The machine is started on the target and the source releases the resources previously owned by the VM

VM Migration modelling

Energy model • Initiation Phase • Transfer Phase • Activation Phase

VM Migration runtime modelling •

Workload types • CPU-BOUND – Computationally-intensive applications • MEMORY-BOUND – Using more than 90% of memory – Memory pages are continuously updated • Dirtying rate: percentage of memory pages dirtied at each iteration

Experimental setup • VM INSTANCES INSTANCE NAME CPU [%] MEMORY [%] DIRTYING RATE [%] VMCPU 100 5 5 VMMEM 100 [5 -95]

Experimental setup EXPERIMENT Source host configuration Target host configuration CPULOADSOURCE Increasing CPU load IDLE up to 125% CPULOADTARGET VMCPU Increasing CPU load up to 125% MEMLOAD-VM VMMEM IDLE MEMLOADSOURCE Increasing CPU load IDLE up to 125% MEMLOADTARGET VMMEM Increasing CPU load up to 125% Migrating VM VMCPU VMMEM [5 -95]% DR VMMEM 95% DR

Measurement infrastructure

Model evaluation (comparison against real measurements) MODEL PIECEWISE ARCHITECTURE ERROR OPTERON 6. 2 % XEON 6. 8% LINEAR (Simgrid) OPTERON 15. 6% XEON 13. 9% CUBIC (Greencloud) OPTERON 23. 6% XEON 20. 4%

Implementation • We implemented our model inside Groud. Sim+DISSECT-CF simulation framework – Groud. Sim is responsible for the workflow execution, while DISSECT-CF is responsible of the infrastructure simulation • We extend DISSECT-CF by – Including power modelling for each component – Adding the VM live migration functionality

Validation • Comparison of real execution traces vs the execution traces obtained with Groud. Sim+DISSECT-CF, simulating the same configuration we used to obtain our measurements. • First figure is the execution on AMD Opteron, the second is on Intel Xeon

Comparison vs Cloud. Simulator Groud. Sim+ DISSECT-CF Architecture Error AMD Opteron 9% Intel Xeon 25. 6% AMD Opteron 45. 9% Intel Xeon 68% Cloud. Sim

Conclusion and future work • We increased accuracy of existing simulators at least by 36. 8% (AMD Opteron) comparing with existing state-ofart simulators • As future work, we plan to: – Further validation of our work with other benchmarks (such as TPC-C and SPEC) – Extension of our model to consider the effects of network and I/O-intensive applications

Thanks for your attention! vincenzo@dps. uibk. ac. at