Challenges towards Elastic Power Management. in Internet Data Center. Introduction. Fast growing of IT power consumption. Cloud Computing -> Internet Data Center(IDCs). The electronic and mechanical systems for power distribution & cooling is the biggest portion of IDC
in Internet Data Center
(CS : such as application organization, load distribution, machine virtualization
PS : such as power distribution, cooling control
It needs coordinations.
They can take advantage of server-level parallelism to scale out in addition to scale up.
They must be easily replicated and migrated at anytime and anywhere.
They can leverage data center level software infrastructure such as MapReduce , Dryad ,EC2 to perform data intensive parallel operations.
Their performances can degrade gracefully when reaching resource limitations.
(Load balancing, VM migration server repurpose.)
If one app’s requirement is low, fulfill it by re-purposing.
Use it to improve the utilization.
It takes information such as service-level agreement , application structures, and environmental conditions , and physical facility constraints from facility and applications designs ; monitors the operation status from application , system , and physical data collected over and across data centers; and makes decisions that affect power provisioning , cooling control , server allocation , service placement , load balancing ,and job priorities.
An important role: to build and refine models to predict performance impacts and risks on resource allocation decisions and to diagnose possible failures.
Such models may in turn become abstractions that designer can use to refine their design so resource utilization can be further optimized.
It may consist multiple sub-layers that are distributed.
Challenge : HOW to organize this layer to perform desired coordination with efficient communication among sub-modules
CPU energy-efficient techniques.
Chip Multi-Processing technology.
Challenge : HOW to group VMs together since hardware resource utilization across VMs are not additive.
Example: two disk IO intensive applications on the same host machine may cause significant throughput degradation due to disk contention.
Challenge: How to integrate techniques at different layers. (DVFS v.s. On/Off, sensitivity of CRAC)
Challenge: lack of consistent abstraction and modularity in computation and physical dynamics.
deal with amount of data , such as VM migration
how to structure the systems ,what data to sense…