Green Latency-aware Data Deployment in Data Centers: Balancing - - PowerPoint PPT Presentation

Green Latency-aware Data Deployment in Data Centers: Balancing Latency, Energy in Networks and Servers Yuqi Fan, Hongli Ding, Donghui Hu

School of Computer and Information Hefei University of Technology

1 School of Computer & Information, Hefei University of Technology, China

August 18, 2014

Model

DC 2 DC 2 DC 3 DC 1 DC 3 DC i

2 School of Computer & Information, Hefei University of Technology, China

Motivation

• Two concerns exist in service provisioning by

data centers

– Users require to experience low latency while Users require to experience low latency while accessing data from the data centers Reduce the power consumed by network transport – Reduce the power consumed by network transport and servers in the data centers

3 School of Computer & Information, Hefei University of Technology, China

Problem

• We tackle the problem of green data deployment in

the data centers, taking into account the three factors

• f latency, energy consumption of the data centers

and the network transport

• The cost of deploying data on a server in a data

The cost of deploying data on a server in a data center integrates the three factors above h f h ffi i i h f i – each factor has a coefficient in the cost function

4 School of Computer & Information, Hefei University of Technology, China

Objective Function

5 School of Computer & Information, Hefei University of Technology, China

• , , and are the weights of the sub-objectives of

h l h i f h d

1



2



3

 the latency, the energy consumption of the data centers and the network transport, respectively.

• rep(dcj , sm, dk) indicates whether data dk is deployed

in server sm in data center dcj .

j

• p(ui | dk) is the probability that a given request is

asking for data dk and it comes from user group ui g

k

g p

i.

• l(ui, dcj) is the latency between user group ui and data

center dcj center dcj.

• rep(dcj, sm) is the indicator whether server sm in data

center dc has been deployed some data

6 School of Computer & Information, Hefei University of Technology, China

center dcj has been deployed some data

• eS(dcj , sm) is the energy consumption of server sm in

d d data center dcj.

• s(dk) is the size of data dk.
• eI (ui, dcj) is the energy required to transport one bit

from data center dcj to user group ui through the Internet Internet.

• PUE(dcj) is the PUE of data center dcj is the power of

sever s in data center dcj sever sm in data center dcj.

• C(sm, dcj) is the capacity of server sm in data center

Dcj.

j

7 School of Computer & Information, Hefei University of Technology, China

GLDD (Green Latency-aware Data Deployment) Deployment)

• When processing each data chunk dk, GLDD

h h i ll h d i h h searches the servers in all the data centers with the least cost to deploy data dk.

• Each server in each data center is checked to obtain

the cost to accommodate data dk on the server if the server has enough capacity.

• The cost of deploying data dk on server sm in data

p y g

k m

center dcj integrates the three factors of the latency, the power consumed by the servers and the network p y transport.

8 School of Computer & Information, Hefei University of Technology, China

9 School of Computer & Information, Hefei University of Technology, China

Result

We evaluate the performance of the algorithm GLDD by comparing GLDD with the algorithm FORTE proposed in SIGCOMM'12. p p

10 School of Computer & Information, Hefei University of Technology, China