POWER BUDGETING FOR VIRTUALIZED DATA CENTERS Harold Lim (Duke - - PowerPoint PPT Presentation

POWER BUDGETING FOR VIRTUALIZED DATA CENTERS

Harold Lim (Duke University) Aman Kansal (Microsoft Research) Jie Liu (Microsoft Research)

Power Concerns in Data Centers

 Consumption costs  Provisioning costs  Cost of supply infrastructure, generators, backup UPSs  Can be higher than consumption cost in large data centers due to

discounted/bulk price on consumption

 Addressed through peak power management

Provisioning Cost Data from James Hamilton

Over-subscription reduces provisioning cost

 Lower allocated capacity => lower provisioning cost (Slight perf hit)  Possible because power can be capped if exceeds [Lefurgy et al.

2003, Femal et. al 2005, Urgaonkar et al. 2009, Wang et al. 2010]

Data Center Power

Rated peak (never reached) Allocated Capacity

Actual power consumption (peak

• f the sum usually lower than

allocated, but can exceed)

Possible peak (sum of server peaks)

Time

Savings from Power Capping

Enter Virtualization

 Existing capping methods fall short

 Servers shared by VMs from different applications: cannot cap

a server or blade cluster in hardware

VM VM … VM VM … VM VM … Rack Server-12 … Server-1j Server-11

Challenge 1: Disconnect Between Physical Layout and Logical Organization of Resources Server

VM1 VM2 Existing Hardware Capping: Unaware of Applications

Server

VM1 VM2 Need: Application Aware Capping

Challenge 1: Disconnect Between Physical Layout and Logical Organization of Resources Server

VM1 VM2

Server

VM1 VM2 Existing Hardware Capping: Unaware of Applications Need: Application Aware Capping

Challenge 2: Multi-dimensional Power Control

100 150 200 250 300 350 400 450 500 550 600 1 2 3 4 5

Performance (TPS) Power (Watt)

DVFS = 100 DVFS = 94 DVFS = 88 DVFS = 82 DVFS = 76 DVFS = 70

Two knobs: DVFS and CPU time cap Different marks are different DVFS levels, multiple marks correspond to different CPU time caps

Perf gap at same power

Challenge 3: Dynamic Power Proportions

 Applications’ input workload volume changes over time

 Proportion among applications changes  Proportion of power among app tiers changes CPU Disk

Front-End Back-End Low Load 50W (CPU Idle) 80W (Disk Spinning, Low IO)

CPU Disk

Front-End Back-End High Load 100W (CPU busy) 90W (Disk Spinning, High IO)

Virtualized Power Shifting (VPS): A Power Budgeting System for Virtualized Infrastructures

Addresses the above three challenges

 Application-aware

 Eg. Interactive apps not affected during capping

 Shifts power dynamically as workloads change

 Distributes power among applications and application

tiers for best performance

 Exploits performance information (if available) and

multiple power knobs

 Selects optimal operating point within power budget

Application-aware Hierarchy

PT(t) Data Center Controller Application Level Controller 1

…

Tier Level Controller 1 Tier Level Controller n Application Level Controller n

…

Tier Level Controller 1 Tier Level Controller n

…

VM VM

…

VM VM

…

VM VM

…

VM VM

…

Papp-1(t) Papp-n(t) Ptier-1(t) Ptier-n(t) Ptier-1(t) Ptier-n(t)

Top Level Controller: Issues

 Determines amount of power for

each application

 Static allocations does not work

 Dynamic workloads and power

usage

 Unused power wasted

 Must compensate for hidden

power increase in shared infrastructure (e.g., cooling load) that are hard to assign to each application

PT(t) Data Center Controller App 1 App n

…

HVAC, Etc.

Top Level Controller: Solution

 Uses feedback (PID) to adapt to dynamic workload and power  Estimates uncontrollable power

 PU(t) = PM(t) – Sum(Pai(t))

 Outputs application power to be allocated

 Papp(t+1) = PM(t) + D(t+1) - PU(t)

Top Level Controller App 1 App m … Data Center Power PT(t) PA1(t) PAm(t) PM(t)

Top Level Controller: Power Split

How is Papp distributed among apps?

 Using Weighted Fair Sharing (WFS)

 Each application has an initial budget

 E.g., 99th percentile of its max power

 In each priority class, allocate power needed to each

app, up to its initial budget

 If not enough power, allocate proportion via WFS

Application Level Controller: Issues

 Determines how much

budget to allocate to each tier

 Prior work: Learn model

• f power ratios among

tiers a-priori. Problems:

 Model changes with

workload

 Depends on the control

knobs used

 Application behavior

may change over time

Tier 1

…

Application

Tier N

Application Level Controller: Solution

 VPS: dynamically tunes power allocations without

relying on learned models

 Observations:

 Application tiers are arranged in a pipeline  Throttling one tier affects other tiers

Tier 1

…

Application Tier N

Application Level Controller (contd.)

 Uses PID control

 Measures total application power usage but only

control one tier

 Automatically achieves right proportion

Controller

Tier 1 (Controlled) Tier n

• Budget

…

u(t) e(t) Pa(t)

Controller:

Tier Level Controller

 Tracks tier power

budget by controlling VM power usage

 Many power control

knobs available

 Use DVFS and VM CPU

time allocation as knobs

PT(t) Tier Level Controller VM 1 VM n

…

 Multiple trade-offs exist w.r.t accuracy, speed,

models needed, app visibility needed

 Study 3 design options

Option 1: Open Loop Control

 Uses power model to convert power budget to

control knob setting

 E.g., PVM=cfreq*ucpu

 Easy and instantaneous  Does not require visibility into application

performance

 But does not compensate for errors

Option 2: PID Control

 Real time measurements to tune power settings:

compensates for error

 Slower (needs time to converge)  Single control knob (no notion of performance

• ptimality)

Controller VM1 VM k

• Tier

Budget … u(t), VM CPU Time e(t) Sum of VM power consumptions

Option 3: Model Predictive Control (MPC)

 Optimizes performance using multiple power control

knobs (DVFS and VM CPU time)

 Uses a cost function that consists of error and performance terms  Solves for the optimal outputs for the next N steps but only apply

the setting for next time step

 Requires application performance measurement  Requires system models that relate control knobs to system

state

Summary of Design Options

Pros Cons

Open Loop Fast Needs power models Higher error PID Low error No performance

• ptimization

Slower MPC Optimizes performance Needs system models Needs performance measurement

Experiments

 VPS controllers run as

network services in root VM on each server

 Controller tuned using

known methods

Physical Server VM

… VM

Controller Service Root VM

 Testbed: 17 Quad core HP Proliant servers (11 host

the apps, 6 generate the workload)

 VMs mixed across the physical servers  VM power measured using Joulemeter, Hardware

power using WattsUp PRO meters

Experiment Workloads

 Applications

 Interactive: StockTrader – open source multi-tiered cluster

web application benchmark

 3 instances, 2 are High priority

 Background: SPEC CPU 2006 benchmark

 Low priority  Use Microsoft data center traces as input to simulate

realistic workloads that vary over time

20 40 60 80 100 Workload (%) Time (s)

Metric: Total Budgeting Error

 Error = excess power consumed above the assigned

budget, normalized by the power budget

0.05 0.1 0.15 0.2 0.25 0.3 0.35 Open Loop PID MPC Physical Hierarchy Overshoot Error (%) VPS

Metrics: Errors within App Hierarchy

 Application power enforcement errors

0.00% 5.00% 10.00% 15.00% 20.00% 25.00% 30.00% Lo-1 Lo-2 Hi-1 Hi-2 Error (%) Open Loop PID MPC

Metric: Power Differentiation

 VPS is designed to respect application priorities and QoS

constraints in a shared infrastructure

 PID and MPC perform appropriate application differentiation

5 10 15 20 25 30 35 40 Open Loop PID MPC Physical Hierarchy Power Reduction (%) Lo-1 Lo-2 Hi-1 Hi-2

Metric: Application Performance

 Performance of (low priority) app that was capped

0.2 0.4 0.6 1000 2000 3000 4000 5000

Response Time (s) Time (s)

PID MPC 2000 3000 4000 5000 1000 2000 3000 4000 5000

Throughput (TPS) Time (s) PID MPC

Conclusions

 VPS: power budgeting system for virtualized data

centers

 Hierarchy of control follows application layout

 Respects application priorities and application VM

boundaries

 Optimizes application performance, given a power

budget

 Dynamically adjusts power proportions  Exploits multiple knobs