Presentation Robayet Nasim (IEEE CloudCom 2016)

OPTIMIZING VIRTUAL MACHINE CONSOLIDATION IN VIRTUALIZED DATACENTERS USING RESOURCE

SENSITIVITY

ROBAYET NASIM, JAVID TAHERI, ANDREAS KASSLER

KARLSTAD UNIVERSITY [email protected]

IEEE CLOUDCOM 2016, LUXEMBOURG, DECEMBER 2016.

ROBAYET NASIM 14-DEC-16

VIRTUALIZED DATACENTERS

VM

Different applications running on virtual machines (VMs) share same physical machine (PM)

Why ? Improved resource utilization

VM

VM

VM

VM

VM

VM

ROBAYET NASIM 14-DEC-16

VM CONSOLIDATION

VM

But? VMs share many resources. Hard to ensure proper isolation. Contention? Specially, when VMs compete for the same resource type.

Why ? VM workload varies over time

VM

VM

VM

VM

VM

VM

VM

VM

Reduce energy consumption

ROBAYET NASIM 14-DEC-16

VM Co-location interference depends on – VM Placement – VMs workload demand – Sensitivity of VMs in relation to different resource types

PERFORMANCE DEGRADATION

How can we improve the VM placement/migration given we have knowledge of VM sensitivity values to different resource types?

ROBAYET NASIM 14-DEC-16

VM MIGRATION/PLACEMENT – SENSITIVITY ?

VM VM

VM

VM

VM sensitive to CPU

VM sensitive to disk

VM

VM

Which migration/placement option should we choose?

ROBAYET NASIM 14-DEC-16

RESOURCE SENSITIVITY AND PERFORMANCE PREDICTION

VMware vSphere

VMware vCenter Server

Manage

VMware vSphere

vmBBProfiler

vmProfiler vmDataAnalyzer

𝑆𝑆𝑆𝑆𝑆𝑆𝑐𝑐𝑆𝑆𝑆𝑆𝑆𝑆𝑚𝑚𝑆𝑆𝑆𝑆𝑆𝑆𝑑𝑑

ProfTable vmLimiter vmDataCollector

vmBBThrPred

vmModeler vmPredictor

�𝑇𝑇𝑇𝑇𝑇(𝐶𝐶,𝑀𝑀,𝐷𝐷)𝑃𝑃𝐷𝐷(𝐶𝐶,𝑀𝑀,𝐷𝐷)

ROBAYET NASIM 14-DEC-16

Design a theoretical optimal interference-aware VM live migration strategy.

Using sensitivity values for the VMs for different resource types in order to minimize co-location performance penalty for a given energy budget.

Validate the proposal using well known applications with various resource signatures – varying from pure CPU/Mem/Disk-intensive to mixed of them.

MAIN CONTRIBUTIONS

ROBAYET NASIM 14-DEC-16

SENSITIVITY AWARE VM CONSOLIDATION MODEL

Power consumption of a PM can be modeled as linear function of resource utilization (such as CPU load, etc.)

Utilization of the VMs allocated to a PM

Budget constraint

Limit on maximum power consumption

Allocated resource to the VMs

Utilization of PMs

Resource demands of old assignments VM migrating towards server VM migrating away

Overbooking factor

ROBAYET NASIM 14-DEC-16

SENSITIVITY AWARE VM CONSOLIDATION MODEL

Sensitivity calculation Objective is to

– Calculate the sum of sensitivity of all the VMs for all resources allocated to a PM after consolidation.

– Minimize variance of total sensitivity for all resources among PMs.

– The limit on energy consumption should not be violated.

Sensitivity of resource i for VM k

Mean of sensitivity for resource i

New VM allocations

ROBAYET NASIM 14-DEC-16

Ensure valid migrations – Migrating within same PM is not

possible, etc.

Other constraints …

SENSITIVITY AWARE VM CONSOLIDATION MODEL

ROBAYET NASIM 14-DEC-16

Matlab is used as problem modeler with IBM CPLEX as solver. – Tune “epgap” parameter to accept relaxed integer solution.

Small scenario to demonstrate model capabilities – 20 PMs, 45 VMs. – 5 selected applications from Phoronix test suite (v5.2.1) (168 available) . – Different levels of resource demands for the same application. – Different allowed overbooking on PMs – 50%, 100%, 150%, 200% .

Solved the model twice for each configuration. – Sen-Aware: Quantitative measures to reflect the sensitivity of the applications. – Sen-Oblivious: Equal sensitivity for all the applications.

EVALUATION

ROBAYET NASIM 14-DEC-16

SELECTED APPLICATIONS

ROBAYET NASIM 14-DEC-16

Case 1: Impact on performance degradation, – Three different levels of resource demands,

Case 2: Impact on fair share of available resources Case 3: Impact of resource demands on throughput.

EVALUATION SCENARIOS

ROBAYET NASIM 14-DEC-16

CASE 1: PERFORMANCE DEGRADATION

Considering Sensitivity value during Migration Helps to reduce Performance degradation

Resource demand -

ROBAYET NASIM 14-DEC-16

CASE 1: PERFORMANCE DEGRADATION

Considering Sensitivity value during Migration Helps to reduce Performance degradation

Resource demand -

ROBAYET NASIM 14-DEC-16

CASE 1: PERFORMANCE DEGRADATION

Considering Sensitivity value during Migration Helps to reduce Performance degradation

Resource demand -

ROBAYET NASIM 14-DEC-16

CASE 2: FAIR ALLOCATION OF RESOURCES

Considering Sensitivity value during Migration Also ensures allocation fairness

Resource demand - , Overbooking - 200%, Energy limit - 2725 W

ROBAYET NASIM 14-DEC-16

CASE 3: RESOURCE DEMANDS VS THROUGHPUT

Amount of Reserve resources for the applications has noticeable impact on their throughput

Resource demand -

ROBAYET NASIM 14-DEC-16

Conclusions – Importance of using sensitivity information of VMs/applications to their

resources for the VM consolidation problem. – Efficient migrations by significantly reducing performance degradation of

VMs especially for higher levels of overbooking. – Not intended for online optimization but can be used to compare the

performance of any heuristic.

Future work – Evaluate how the accuracy of the sensitivity estimation affect the overall performance. – Design fast heuristic while taking into account the sensitivity values and integrate them into online optimization tool such as OpenStack Watcher.

CONCLUSIONS AND FUTURE WORK

ROBAYET NASIM 14-DEC-16