Project
|
|
This page describes the project for Simulation.
Electricity costs have become the largest cost in operating a data center
exceeding hardware, software, and personnel costs. The energy use
of data centers is also a major environmental concern. A 2007 Report to Congress
on data center energy use is here.
Not surprisingly, there is significant ongoing research and development in
finding ways to reduce energy use of data centers. For your semester project you
will evaluate the performance of a hybrid server approach where a low-power
and low-performance proxy server covers for a high-power and high performance
master server during periods of low request rates. The master and proxy servers
are (content-wise) identical mirrors of each other. Such a hybrid server
architecture has significant potential to reduce energy costs because workload
to data centers is highly variable (for example, by time of day). A paper
describing early results in our work on hybrid web server is under review for
the 2011 IEEE Conference on Local Computer Networks and is
here.
A YouTube video describing this research is
here.
The System
The system you will model consists of a master server, proxy server, request
director, and a source of requests to the proxy-master hybrid system. A picture
of the system is here (PPT). The request
director is the heart of the system - this unit has the ability to wake-up the
proxy or master server (and put the other server to sleep) and then route all
requests to the awake server. All requests (jobs) arrive to the request
director. Requests sent to a server are queued and serviced in FCFS order.
The key question is, what policy should the request director use to wake-up (and
put to sleep) a server? In other words, when should the request director wake-up
one server and put the other to sleep? Clearly, when the request rate is "high"
then the master should be awake, when the request rate is "low" then the master
should be asleep (saving energy). We will make the following assumptions
regarding the hybrid server system:
- It takes zero time for the request director to route a request to a server.
- It takes zero time to wake-up or put to sleep a server.
- Only one server, master or proxy, can be awake at any given time (note that
this is different from the system described in the above LCN paper).
- Once a server has been woken-up it must be kept on for a fixed period of
time before it can be put to sleep. Nominally, we will assume that 5 minutes is
the minimum time that the master server can be awake and 1 minute is the minimum
time that the proxy server can be awake.
- The request director can keep statistics on arrival rate of requests; it
has no knowledge of server or proxy state (e.g., number of requests queued,
response time, etc.) other than is it awake or asleep.
- The master server power draw is 200 W and the proxy server power draw is
5 W.
- Electricity costs $0.10 per kWh.
- The proxy computer costs $100.
The Workload and Key Factors
The work load you will use is a trace from a realworld corporate server. The
trace consists of timestamps for one month - each time stamp indicating a
request received by the corporate server. The time resolution is 1 second.
The trace is trace.txt (download as a zip file).
To simplify our model a little we will assume that all requests take an
equal amount of time to process (that is, service time). We will assume
that:
- The service time for handling a request on the master is 1 millisecond.
- The service time for handling a request on the proxy is 10 milliseconds.
Performance Measures of Interest
You are interested in studying the trade-off between energy savings (that is,
total sleep time of the master server) and request response time. Clearly, if
you keep the master server awake always, the energy savings will be minimum
and the response time minimized. If you keep the proxy server awake always,
the energy savings savings will be maximized and the response time maximized.
We will deem that a response time of over 100 milliseconds is "too high".
You should determine the percentage of requests that experience a too high
response time. A small percentage (say, 1% to 5%) may be deemed acceptable as
a trade-off to energy savings.
You must also determine the payback time for the extra hardware - the proxy
computer.
Project Deliverables
Each team of two students (students may work individually if they wish) will
deliver the following:
- A maximum five page paper written in IEEE conference paper format and
style describing the problem, system, questions to be answered, simulation model
and workload, model validation, experiments conducted, experiment results,
related work, conclusions, and list of references. The final paper must be
submitted as a PDF file.
- The source code for the simulation model in buildable form (so that the
results can be reproduced).
Project Grading
Project grading is as follows:
- Design of model -- 10 points
- Validation of model -- 5 points
- Design of prediction policies -- 15 points
- Design and execution of experiments -- 10 points
- Correctness of experiment results ----- 40 points
- Conclusions including payback period -- 10 points
- Related work and references -- 10 points
- Organization, grammar, format, and neatness matter. Up to 30 points can be
deducted for weaknesses in these areas.
- Clean code matters. Up to 20 points can be deducted for poorly documented
source code.
Up to 20 extra credit points are possible if the student develops a particularly
novel or sophisticated policy for determining when to switch between the master
and proxy. Such a policy should show better performance than any "obvious"
simple policy (and this better property needs to be shown as such). It is highly
recommended that the student consult with the instructor before attempting this
extra credit.
Notes
Some notes for the projects are:
- There are two key cases that serve as the upper and lower bounds for
performance and energy savings that all papers must give results for.
These are the case of 1) all requests served by the master server, and 2) all
requests served by the proxy server.
- How to determine if your prediction policy is good? You should implement
an oracle policy to serve as a best possible upper bound.
- A related work section is key to any paper. Just googling Hybrid Server
and dumping what you find in the first screen in the paper probably won't do
it. Hint: use scholar.google.com to search for "hybrid datacenters".
- The IEEE conference paper template for Word is
here. Two example papers are
here and
here.
- When in doubt about style (e.g., correct format for references), the IEEE-CS Style
Guide is the master reference.
- The coding style guidelines you must follow are
here.
|
Last update on June 27, 2011
|