Enabling Datacenter Servers to Scale Out Economically and Sustainably Chao Li*, Yang Hu*, Ruijin Zhou*, Ming Liu*, Longjun Liu§, Jingling Yuan†, Tao Li* * Intelligent Design of Efficient Architectures Laboratory (IDEAL), University of Florida § Xi’an Jiaotong University, China;; † Wuhan University of Technology, China

The ‘scale-out’ issue in datacenters o Server consolidation limits workload performance

through SW based or HW based control knobs. Conventional centralized power provisioning scheme does not scale well

Scale out power-constrained data centers o Datacenters are power-constrained

o Datacenters are carbon-constrained

Scale-out model o Distributed incremental integration

Modular power sources o Distributed energy storage system

o Solar power module with micro-inverters

Power control hub (PCH) o A prototype that integrates battery charger,

inverter, power source switch, control devices(PLC&HMI)

Dynamic energy source switching o Autonomous mode

o Coordinated mode

1. Chao Li, Wangyuan Zhang, Chang-Burm Cho, Tao Li, SolarCore: Solar Energy Driven Multi-core Architecture Power Management, in Proceedings of the 17th International Symposium on High-Performance Computer Architecture (HPCA), February 2011

2. Chao Li, Amer Qouneh, Tao Li, iSwitch: Coordinating and Optimizing Renewable Energy Powered Server Clusters, International Symposium on Computer Architecture (ISCA), June 2012

Introduction Technical Approach

Experimental Results

References

System Implementation

Fig. 1: Battery discharging profile and the renewable power supply and server power demand traces (workload: Nutch, page indexing)

Fig. 2: The estimated battery lifetime Fig. 3: The average battery backup capacity

Fine-grained scale out power infrastructure o Oasis leverages modular renewable power

integration and distributed battery architecture to provide flexible power capacity increments

o Could reduce CapEx by 25%

Prototyped research platform o Power supply monitoring

o Communication gateway

o Power supply switching

Optimization algorithm o Oasis could reduce workload execution delay to

1%, extend battery lifetime by over 50%, increase battery autonomy time by 1.9X

Battery management o Optimize battery usage effectiveness

Renewable power prediction o Minimize control overhead

Feedback control on server system o Improve cloud server performance

o Improve robustness

Contributions

Future Work

Node architecture System prototype

Rack Power Strip

Cluster-Level PowerManagement Agent

Inverter

PLCHMI

MPPT

SensorSensor

Switch Panel

Network SwitchEthernet

ModBus

Charger

L N G

Serv

ers

Pow

er C

ontr

ol H

ub

Utility PowerBatteryStatus ← Low

V_battery < V_discharge

N

V_battery > V_charge

V_battery < V_deplete

BatteryStatus ← High

BatteryStatus ← None

SolarStatus ←TsolarOutput > THLD

N

UtilityStatus ←TpduOutput > THLD

N

N

Y

Y

Y

UtilityStatus ← F

SolarStatus ← F

Y

Y

SolarSwitch ← ON

BatteryStatus = High

YSolarOutput = TRUE

Wait 5 sec

N

UtilitySwitch ← OFF

UtilityOutput = TRUE

UtilitySwitch ← ON

Wait 5 sec

SolarSwitch ← OFF

Send Alert YN

YN

(a) SupplySense Module

(b) SupplySwitch Module

0%

20%

40%

60%

80%

Consolidate Server

Upgrade Equipment

Deploy Container

Build New Datacenter

Lease Colocation

Move to Cloud

“How will your organization handle increased capacity demand over the next 12-18 months”

Power Control Hub

Distributed Battery

Grid-Dependent Racks

PDU PDU

ATSUtility Backup

Support Scale-out Servers

Centralized UPS

Offer Partial Carbon Reduction

Micro-Inverters

HMI

PLC

(ModBus TCP/Server)(ModBus TCP/Client)

EthernetRS-232/485

Server OS

Workload Workload

Power Management Agent

Workload

Server System (The Load) Power Control Hub

Oasis power provisioning architecture

Power management agent and the data communication scheme inside of Oasis node.

Battery Discharge Profile

0 50 100 150 200 250 300 350 400 450 5000

100

200

300

Time (minutes)

Pow

er (W

)

Renewable Power Server Power

0%

100% Battery Discharge Profile

0 50 100 150 200 250 300 350 400 450 5000

100

200

300

Time (minutes)

Pow

er (W

)

Renewable Power Server Power

0%

100%

(a) High Renewable Power Variability (b) Low Renewable Power Variability

0

5

10

15

20

25

Lifet

ime

(Yea

rs)

Oasis-B

Oasis-L

Ozone

0%

20%

40%

60%

80%

100%

Back

up C

apac

ity

Oasis-B

Oasis-L

Ozone