Transcript SISO PIDF Controller in an Energy-efficient Multi

Energy-efficient Multi-tier
Web Server Clusters
Daniel Mosse
Introduction
Practical use of power control of multitier clusters
 Application: e-commerce or other
datacenter usages
 Goal: meet deadlines AND minimize
energy
 Open software and commodity hardward

Cluster Model
Power Model


CPUs can vary frequency and voltage
The strategy is: if there is time, slow down
processor (and procesing)
 Real-time tasks have deadlines (that is, tasks
must finish by a certain time)
 How much, then, can we slow down tasks?
Static speed adjustment
Assumption: all tasks have the same deadline.
CPU speed
WCET
deadline
Smax
Smin
time
Smax
Smin
time
Select the speed based on
worst-case execution time,WCET, and deadline
Workload Generation





TPC-W Benchmark for e-commerce (Transaction
Processing Performance Council)
Multiple on-line browser sessions
Dynamic page generation and database access
Authentication through SSL
Real-time properties
Measuring end-to-end time delay
Client
Front-end
req.php
Server
node 1
req.php?r=123
Server
node 2
Server
node 3
Img1.jpg?123
Img2.gif?123
Img3.jpg?123
img2.gif?123
img3.jpg?123
WIRT
img1.jpg?123
Using a Controller

Main metric: Quality of Service
 Web interaction

Web interaction response time (WIRT)

Controller metric:Tardiness of a web interaction

Tardiness show more precise information about
the completion of requests and resource usage
Control servers on-off also!

If there are N machines, and we can put some
machines to sleep, and turn some machines off





How many machines are put to sleep?
How many machines are turned off?
When do we change state of machines?
For machines that are on, how to set speed?
Multi-core systems:


Heterogeneous vs homogeneous cores
Change voltage for each core? What about shared
memory?