[httperf] Project to enhance httperf. (Multiple clients and Session emulation)

Martin Arlitt arlitt at hpl.hp.com
Mon Nov 15 07:05:44 PST 2004


Pralhad

> I belong to a group of 4 students in India, who are doing a project to enhance the features of httperf. We wish to do the following:

is Vishwanath part of this group?

> 1) Implement httperf over multiple clients.

I think Ian has already done this, as he mentioned in his message.

2) Sessoin emulation i.e. Incorporate user behaviour in the requests made to the server. This will involve multiple URL stressing and implementing think time and retries in the sessions.

it may be useful to think of the broader picture here (i.e., Web server
benchmarking in general), and determine what are some of the (remaining)
obstacles to realistic benchmarking.  you may also want to consider a
number of scenarios.

for example, consider a person who is implementing or enhancing a new Web
server (like a person who contributes to the Apache project).  in this
situation the person would like to generate a realistic workload that they
expect would be representative of many Web sites, since they intend for
their enhancements to be widely adopted.  in this case they might want to
use analytic trace generation.  this approach is used by SURGE
(http://www.cs.bu.edu/faculty/crovella/links.html).  in this case their
isn't a lot of effort required to set up the workload generation and
specification, which is desirable, since the person wants to focus on the
Web server enhancements.

one of the advantages httperf has over SURGE is that httperf is an
open-loop workload generator.  this feature allows the tester to
examine the server under overload situations. however, httperf lacks the
User Equivalents of SURGE.  this makes it more challenging to specify a
workload that would be considered realistic and representative of Web
server workloads in general.

a second scenario is where an adminstrator wants to test various features
of the site they operate.  in this case, they may want to use trace-based
workload generation, so that the workload is representative of their own
site (and since they will have access logs from their server).  while
httperf provides the ability to specify sessions, what I believe is
missing is a tool to extract the important characteristics from the trace
and create the input files and parameter settings for httperf.

there are likely other common scenarios, so I would suggest you think
about this a bit before deciding how to proceed.  I can discuss this
further with you if you'd like.

> We would be really very happy to recieve any guidance on the above mentioned points. If any of the above objectives have already been achieved please let us know. We will be pleased to make further improvements if any are required.

how much time do you have to decide on and complete the work?

> Also we would like to know if all the features of httperf are currently working perfectly or any improvements are desired, especially in the area of Session handling.

there aren't a lot of problems that I am aware of.  one that you should be
aware of before trying to use the sessions feature is

http://www.hpl.hp.com/hosted/linux/mail-archives/httperf/2004-February/000151.html

thanks

Martin


More information about the httperf mailing list