OPTIMAL SERVER PROVISIONING AND FREQUENCY ADJUSTMENT IN SERVER CLUSTERS Presented by: Xinying Zheng 09/13/2010 1 XINYING ZHENG, YU CAI MICHIGAN TECHNOLOGICAL.

1

OPTIMAL SERVER PROVISIONING AND FREQUENCY ADJUSTMENT

IN SERVER CLUSTERS

Presented by: Xinying Zheng09/13/2010

XINYING ZHENG, YU CAIMICHIGAN TECHNOLOGICAL UNIVERSITY

2

Outline

Introduction Related Work Optimization problem formulation

Single class Multiple classes

Overhead Analysis: DCP model Performance Evaluation Conclusion and Future Work

3

Motivation

The increased data centers and cluster systems consume significant amount of energy.

4

Motivation

The power consumption of enterprise data centers in the U.S. doubled between 2000 and 2005. And will likely triple in the next few years.

Servers consume 0.5 percent of the worlds total electricity usage, this number will increase to 2 percent by 2020.

5

Benefits of Greening

6

Processor

Memory

DVS( Dynamic voltage scaling)

Feedback Control

DTM( Dynamic thermal management)Single Server

Storage and Database Servers

Web and application Servers

Non-data Movement

Data Movement

DV/FS

Feedback Control

VOVF

DTM

Virtualization

MemoryNetwork Techniques

Discs

Performance level

DVC

Economic method

Wireless sensor networks

Computer networks

Request-response service

Long-live connected service

Server Cluster

7

System modeling

• Request timePerformance

• Cubic power model

Power consumpti

on

8

Syetem Assumption

All servers in the cluster are identical

nodes.

Each server has two modes: active and

inactive.

Operate at a number of discrete

frequencies.

All the incoming requests are CPU

bounded.

9

Performance metric modeling Incoming request follows a heavy-tailed

bounded Pareto distribution.

If we define a function: Average job size:

(1)

(2)

(4)

(5)

(3)

10

Request time in single server

Server processing capacity: c

Packets inter-arrival time follows exponential distribution with a mean of 1/λ.

According to Pollaczek-Khinchin formula, the average waiting time is :

Request time:

(6)

(8)

(7)

(9)

11

Extend to server cluster

Extend to the server-cluster mode. Using Round-Robin dispatching policy, the arrival process at each server in the cluster has rate .

Processing capacity is proportional to frequency.

Request time:

/ m

(10)

(11)

12

Power consumption modeling Power-to-frequency relationship. Linear model. Cubic model:

System power consumption:

(12)

(13)

13

Optimization problem formulation Minimizing total power consumption. Request time threshold. Mechanism:

VOVF: vary-on, vary-off DFS: dynamic frequency scaling.

14

Optimization problem formulation (single class)

Single class:

Computation complexity is O(NM). Complexity can be reduced to O(NM).

applying a coordinated voltage scaling.

(14)

15

Optimization problem formulation (Multiple classes)

Assuming incoming requests are classified into N classes.

The ratio of average request time between class i and j is fixed to the ratio of the corresponding differentiation parameters:

We assume class 1 is the “highest class” and set:

E[Ri]

E[Rj]

i

j

(15)

0 1

2L

w

16

System model of multiple classes

17

Optimization problem formulation (Multiple classes)

Multiple classes:

Different class receive different performance.

(16)

18

Overhead Analysis

Server transfers from inacitve to active mode.

Transition time influence the performance.

Double Control Periods(DCP) model.

Double control periods

19

Overhead Analysis

20

Simulation

Package

generator

•incoming request

•Inter arrival time between package

Load

dispatche

r

•Caculate the number of active servers according to workload.

•Dispatch incoming jobs to active server.

Number of servers

•Waiting queue.

•Excute the jobs in FIFO discipline.

21

Evaluation (single class)

Request time comparison between OP model and DCP model

Power consumption comparison between OP model and DCP model

22

Evaluation (multiple classes)

Request time comparison between OP model and DCP model

Power consumption comparison between OP model and DCP model

23

Evaluation(real workload single class)

Request time comparison between OP model and DCP model

Power consumption comparison between OP model and DCP model

24

Evaluation(real workload multiple classes)

Request time comparison between OP model and DCP model

Power consumption comparison between OP model and DCP model

25

Contributions

Optimization model for power reduction in server clusters.

Single class and multiple classes. Double control periods model to

compensate the transition overhead. Evaluate our models in real workload

data trace.

26

Future work

Effect of dispatching strategy. Transition overhead of frequency

adjustment. heterogeneity in data centers. Apply our model to the real Internet web

servers in the future.

Questions

Thanks for your attention