Energy Efficient Scheduling in IaaS Cloud Mehdi Sheikhalishahi University of Calabria Supervisor:...

Post on 27-Mar-2015

213 views 0 download

Tags:

transcript

Energy Efficient Scheduling in IaaS Cloud

Mehdi Sheikhalishahi

University of Calabria

Supervisor:Prof. Lucio Grandinetti 

OGF 28Munich, 15-19th March

Outline

Scheduling Algorithms in Computing Systems

Energy Efficient Scheduling

2/11

Backfilling (Space sharing): User

Estimated execution time Aggressive, Conservative, Probabilistic, Lookahead Provider

The number of reservations The order of queued jobs (FCFS, priority) The amount of lookahead into the queue

Gang (Time sharing by time-slice) Parallel (MPI) App characteristics: IO-intensive, Compute-intensive, memory-

intensive Context switch, Memory pressure, addi. Swap space pressure

Scheduling Algorithms

3/11

Job/Scheduling parameters and QoS metrics

4/11

5/11

Job/Scheduling parameters and QoS metrics

6/11

•As information

•Hybrid cloud operation

•Scheduling over a number of clouds

Green ComputingClim

ate

chan

ge

Climat

e ch

ange

labellabel

global warmingglobal warming

Enabling technologies (energy efficient)Enabling technologies (energy efficient)

VirtualizationVirtualizationCloud SchedulerCloud Scheduler

Policy, reordering, adjusting frequencyPolicy, reordering, adjusting frequency

ElectricityElectricityHeatHeat carboncarbon

Multicore (DVFS,cpuidle)Multicore (DVFS,cpuidle)

7/11

Energy Efficient Scheduling

Energy Consumption=Electricity for power+cooling+etc Modern processors

Dynamic Voltage and Frequency Scaling (DVFS) Cpufreq

Performance state P-states={(f0,v0,h0), ... ,(fn,vn,hn)}

f(i): frequencyv(i): the voltageh(i): heat generated

Cpuidle Power state C-states={(v0,wl0,h0), ... ,(vn,wln,hn)}

v(i): the voltagewl(i): wakeup latencyh(i): heat generated

Random WorkloadRandom Workload

Short and long jobsShort and long jobs

Different sizesDifferent sizes

Inaccurate estimated execution timeInaccurate estimated execution time

Scheduling metricsScheduling metricsDVFS

fragmentation and utilizationfragmentation and utilization

8/11

Energy Efficient Scheduling

Energy Consumption Model

Two different operating points a(va,sa) and b(vb,sb),

Two types of scheduling(Performance,Energy) XEN hypervisor

Interfaces and governor (xenpm) Extend running time to fill out the availability window by reducing

speed of processor

9/11

Job/Energy Efficient Scheduling parameters

10/11

Virtualization technology: Job/Scheduling parameters

11/11

Thanks

?

12/12