vGreen: A System for Energy

crashclappergapSoftware and s/w Development

Dec 13, 2013 (3 years and 10 months ago)

64 views

vGreen
: A System for Energy
Efficient Manager in Virtualized
Environments

G.
Dhiman
, G
Marchetti
, T
Rosing

ISLPED 2009

vGreen


Multi
-
tiered software system for energy
efficient computing and management in
virtualized environments.


Captures power and performance
characteristics of virtual machines and
develops policies for energy efficient VM
scheduling.


Performance and system level energy savings
of 20% and 15%

Importance


Power Consumption critical


because it impacts deployment (peak power delivery)


Affects operational costs (power supply, cooling)


Current work treats overall CPU utilization of PM
and its VM as indicator for power consumption
and resource utilization


Characteristics of co located VMs causes variation
in power consumption at similar CPU utilization
levels.

Solution technique


vGreen


Understand and exploit relationship between
architectural characteristics of VM and its
performance and power consumption.


Architectural characteristics comprise of
instructions per cycle, memory access


Based on client server model


Vgserv and vgnodes



vgserv and vgnodes


Vgserv


Centralized server


Performs management decisions like scheduling and
DVFS of VMs across PMs


Places VMs across
vgnodes

to improve overall
performance


Vgnodes


Physical Machines where VMs located


Perform online characterization of the VMs running
on them and updates
vgserv



Principle and methodology


Nature of workload executed in each VM determines
the power profile and performance of the VM, and
thereby its energy consumption.


VMs with different or same characteristics co
-
located
in same VM


Characteristics refer to CPU and memory utilization


Two contrasting benchmarks mcf and perl used to
implement heterogeneous characteristics


eon and
mcf


mcf


High Memory Accesses per cycle (MPC)


Results in increased cache conflict rate for
multiple instances


Increased execution time


eon


Has high Instructions per cycle but low MPC


Results in higher utilization of CPU resource


Comparison of
mcf

and eon

Conclusion from results


Co
-
scheduling VMs with similar characteristics not
beneficial from energy efficiency and power
consumption point of view.


mcf

contributes to higher system energy
consumption because of its longer running time.


eon contributes to power imbalance as it consumes
more power


Running VMs with
mcf

and eon on both PMs result in
high performance improvement and energy savings
upto

20%


Power Management in Virtualized

Hierarchical Metrics

Explanation


vgpolicy decisions based on value of different
metrics, namely MPC, IPC and utilization of different
VMs


These metrics received as updates from vgnodes.


Metrics evaluated and updated dynamically



Continued…


vgxen estimates aggregate metrics (vMPC,
vIPC, vutil) for each VM by adding up metrics
of constituent VCPU and stores it and exports
it to vgpolicy through vgdom vgnode.


vgdom acts as interface for vgnode to vgserv
and registers vgnode with vgserv.

MPC balance algorithm

Explanation


Checks if
nMPC

of n1 greater than threshold MPC.


Return if small otherwise find VM with minimum
vMPC

in n1 and migrate it to
vgnode

with lower
nMPC

for better balance.


But migration should not result in the
nMPC

of new
node exceeding threshold MPC.


Same procedure for IPC.


Utilization is balanced to ensure no overcommitted
or underutilized node exists.


VM consolidation of low utilization VM to idle VM


DVFS


Vgpolicy

issues command to scale v
-
f setting if
it is more energy efficient than VM migration.


Can be required if heterogeneous VMs are
absent.


Exploit characteristics of workload to find v
-
f
setting that is best suited.


mcf

and eon run at 90% CPU utilization levels

Different frequency levels

MPC, IPC, DVFS


MPC highest priority


Memory bottleneck impacts performance and
energy efficiency


IPC next


Balanced power consumption, results in uniform
thermal profile and decreases cooling cost.


Utilization for fair distribution of workload
.


DVFS when no benefits obtained from VM
scheduling

Average Weighted Speedup


Average Weighted speedup




T
e+i

= time of execution of
VM
i

with E+


T
vgreeni

= time of execution of
VM
i

with
vGreen


T
alonei

= time of execution of
VM
i

running
alone on
VM
i



Mixed
vs

Same VM placement

Weighted Speedup and Energy Savings

Power Consumption

Conclusion


vGreen has negligible runtime overhead


Workload characterization achieves better
performance and energy efficiency


Reduces power consumption variance between two
vgnodes by 80%