This document answers frequently-asked questions about WebStone.
If you have comments about this document, please forward them to the author.
This is a list of answers to Frequently Asked Questions (FAQ) about WebStone. The latest copy is always available at http://www.sgi.com/Products/WebFORCE/WebStone/ and via the WebStone mailing list. The FAQ is periodically posted to the WebStone mailing list, and to the USENET newsgroup comp.benchmarks.
WebStone 2.0 is a rewrite of the WebStone 1.1 code. Significant changes have been made both in the code and in the fileset and run rules. Many bugs were eliminated, support for other platforms has been included and many new features have been added. The WebStone 1.1 and WebStone 2.0 numbers cannot be compared, since so much has changed. In general, WebStone 1.1 will give higher connections/second values, but lower throughput numbers than WebStone 2.0.
Absolutely NOT! WebStone 1.1 numbers are based on a different fileset, as well as an older version of the benchmarking software. The WebStone 1.1 fileset was based on a fileset with a smaller average filesize, so that the number of connections per second will tend to be higher (all things being equal). The WebStone 2.0 fileset is based on observations of several real world sites, and the distribution of the filesizes found there. This fileset is also similar to the fileset chosen by the SPEC committee for their benchmark.
While it is possible to convert the 1.1 fileset to a 2.0 format and then test it, the resulting numbers will not be the same, because the underlying software used to perform the testing has changed. WebStone 1.1 was also heavily abused because of the lack of run rules, and reporting rules. It is recommended that everyone move to WebStone 2.0.
WebStone is a highly-configurable client-server benchmark for HTTP servers.
The original WebStone benchmark was released in March, 1995. The original white paper describing this benchmark is available from http://www.sgi.com/Products/WebFORCE/WebStone/.
WebStone is not a proprietary benchmark - it is an open benchmark. The source code is freely available, and anyone can examine it. By design, WebStone does not unfairly favor SGI, Netscape, or any other company - it is simply a performance measurement tool.
A SPEC SFS working group is presently adapting SPEC SFS to Web server benchmarking. SGI's WebStone team is part of this working group, and we support fully the effort. WebStone is available to fulfill the immediate Web benchmarking needs - not to confuse the public.
Basically, if you like WebStone, use it. When SPEC releases Webperf, check it out.
WebStone makes a user-configurable number of HTTP 1.0 GET requests for specific pages on a Web server. Any Web server can be tested, and any HTML content can be used.
WebStone measures the throughput and latency of each HTTP transfer. By default, only statistical data are returned, but the user may optionally request data for each and every transaction. WebStone also reports transaction failures, which translate into those little "Connection Refused" alerts in the real world.
WebStone 2.0 includes support for testing proxy servers, as well as more flexible handling of URL's that enable WebStone to test a wide variety of content types. The code has also been significantly rewritten so that it is more robust and portable.
WebStone does not yet do any of the following (listed roughly in order of planned implementation):
If you have additional requests for WebStone functionality, contact the WebStone mailing list.
The latest copy of WebStone, and of this FAQ, is available at http://www.sgi.com/Products/WebFORCE/WebStone
WebStone includes a README file which may answer some of your questions. However, here's a brief overview.
testbed
To try the GUI, make sure you have a Web browser, and run ./webstone -gui from the WebStone base directory. You don't need to hand-edit the testbed file anymore, but you still need to edit filelist if you want to change the workload. This may not be necessary, since we've
distributed two real-world workload models with WebStone.
These are the stepts to follow to run the GUI
./configure
./webstone -gui
If the GUI appears to hang, you can kill stray WebStone processes with ./webstone -kill
Your test bed should include, at minimum, two machines and a network. The first machine is your Web server - it can be any HTTP 1.0-compliant server. As far as WebStone is concerned, it's a black box.
You'll also need a webmaster and one or more webclients. These should be Unix hosts, since WebStone hasn't been tested on any non-Unix operating systems (feel free to port it, if you like). The webmaster and the webclient may be the same machine, if desired: we've run up to 120 webclients and the webmaster on a single 32MB Indy.
You must establish a trust relationship between your webmaster and webclients.
Each webclient must be set up so that the webmaster can use rexec to execute the WebStone on the client. This can be done with a guest account.
It's also helpful if root can rexec and rcp to the webclients, and even to the web server. This requires editing the /.rhosts and /etc/host.equiv files. Here's an example:
/.rhosts (on each webclient)
webmaster root
/etc/hosts/equiv (on each webclient)
webmaster
To make best use of WebStone, your webmaster should be equipped with a C compiler, Perl, awk, and a Web browser. A data analysis program such as GnuPlot may also come in handy.
Connect the webclients, the webmaster, and the web server to a common network. To check your setup, load a browser on one of the webclients, and make sure it can connect to the Web server.
Copy the WebStone distribution onto your webmaster. If your webmaster isn't an SGI IRIX 5.3 machine, you'll have to make the binaries. Type make from the WebStone directory - this creates the following binaries:
webmaster webclient
Common porting errors
src/Makefile.
LIBS = -lsocket -lnsl in src/Makefile.
rexec in webmaster.c
If you encounter other errors, please contact the WebStone mailing list.
Type make install to put the binaries in the bin directory.
When you run WebStone, the distribute script automatically copies the webclient binary to the other client systems. If you're running diverse clients (e.g.,
a couple Suns, a couple BSD hosts), you'll want to comment the distribute script out of bin/runbench, and distribute host-specific versions of webclient by hand.
If you use the webstone script to automate WebStone, you'll want to edit the conf/testbed script. The testbed script contains several configurable parameters that WebStone relies on.
Here is an example:
### BENCHMARK PARAMETERS -- EDIT THESE AS REQUIRED ITERATIONS="3" MINCLIENTS="8" MAXCLIENTS="128" CLIENTINCR="8" TIMEPERRUN="30" ### SERVER PARAMETERS -- EDIT AS REQUIRED #PROXY= SERVER="www" PORTNO=80 SERVERINFO=hinv OSTUNINGFILES="/var/sysgen/master.d/bsd" WEBSERVERDIR="/usr/ns-home" WEBDOCDIR="$WEBSERVERDIR/docs" WEBSERVERTUNINGFILES="$WEBSERVERDIR/httpd-80/config/magnus.conf $WEBSERVERDIR/httpd-80/config/obj.conf" # WE NEED AN ACCOUNT WITH A FIXED PASSWORD, SO WE CAN REXEC # THE WEBSTONE CLIENTS CLIENTS="webstone1 webstone2 webstone3 webstone4 webstone5" CLIENTACCOUNT=guest CLIENTPASSWORD=guest CLIENTINFO=hinv TMPDIR=/tmp
Briefly, the first set of parameters means that the WebStone benchmark will run from 8 clients to 128 clients, in increments of 8. Each increment will run for 30 minutes, and the whole test will be repeated three times. This test suite would take roughly 24 hours to complete.
Why multiple iterations? The WebStone benchmark is a stochastic process so there will be variation from run to run, especially if your test file sets have large files or if you approach overloading the server. 3 iterations is about the minimum you should run just to see if there is variation and to gauge the amount of variation. the TIMEPERRUN needs to be long enough to establish a steady state and allow it to dominate the run. 30 minutes seems to be enough if the sizes of the files are small. You may want to run the benchmark longer per run to minimize variation if the files are large.
The second set of parameters means that we will test a server called "www"
at port 80 (note that the port number may be changed to accomodate proxy
servers or multiple servers on the same host). We will use four clients.
Also, we specify the location of a system tuning file (on Sun Solaris, one
could use /etc/system), and web server tuning files (specified for Netscape).
These files will be copied into the runs subdirectories for later reference.
Finally, we specify the WebStone account on the clients. Here, we use the guest account, with a fixed password: guest.
The basic WebStone tests expect a set of files to reside on the server to be retrieved by the webstone client programs. The file list tells WebStone which files to retrieve.
It's possible to use an arbitrary set of fixed-length files for WebStone. Although these files have the .html extension, they are used to represent files of many types. Basically we treat "bits-as-bits". You can use the programs in the genfileset subdirectory to create the needed set of files, and copy them onto your server:
./webstone -genfiles
The sample file list shipped with WebStone uses the files created by genfiles:
# Sample filelist, abstracted from access logs
/file500.html 350 #500
/file5k.html 500 #5125
/file50k.html 140 #51250
/file500k.html 9 #512500
/file5m.html 1 #5248000
This filelist consists of 5 different files. The number following the filename is the weight of this file in the distribution. All the weights are summed together and the frequency of each file is the weight of that file over the total weights.
For example, in this fileset the weights add up to 1000. So the the file500k.html page will occur 350 out of 1000 times, and the file5m.html will occur once every 1000 pages.
Note that the URI should be changed to a full URI when testing proxy servers, for example, if the proxy server is called proxy, but the actual server which stores the file is called seltzer1, you could use the following filelist:
#Sample filelist, abstracted from access logs
http://seltzer1.sgi.com/file500.html 350 #500
http://seltzer1.sgi.com/file5k.html 500 #5125
http://seltzer1.sgi.com/file50k.html 140 #51250
http://seltzer1.sgi.com/file500k.html 9 #512500
http://seltzer1.sgi.com/file5m.html 1 #5248000
This URI is the one which is passed to the proxy server, which in turn uses it to fetch the file from seltzer1.sgi.com. Notice that the particular files and the distribution are identical to the previous filelist. The other change which would need to be made for testing proxy servers is to have an entry "PROXY=proxy" in the testbed file and to specify the port where the proxy server listens for requests.
Wherever possible, use the same pages for WebStone that you will use in the real world. This means that you'll have a harder time comparing your results with published results, but your results will more accurately reflect your situation.
type ./webstone
The results of each run will be saved in a directory called runs. Note that the runbench script attempts to collect configuration information about your client and server configurations such as netstat results. You may see some error messages if your clients don't have netstat or other utilities.
The WebStone summary statistics generated by webmaster are saved by runbench in a date stamped subdirectory of the runs directory in the current directory similar to:
runs/950804_2304/run
The script wscollect is provided as a tool for collected the results of all of the runs and generating a tab delimited file with all of the results. This file can be read into a spreadsheet or read by other analysis programs.
wscollect runs > runs.tabs
An additional script called tabs2html will take a tab delimited file and produce an HTML 3.0 style table of the results:
tabs2html runs.tabs > runs.html
It's fairly common for the Web server under test to run out of swap space. As a rule of thumb, make sure that you have swap space equal to the number of server processes times the size of the largest test file.
For instance, if you're testing a 10MB file on a Netscape server with 64 processes, you'll need to have at least 640MB of swap space. N.B.: On SGI IRIX 5.x, you can substitute large amounts of virtual swap space, since Netscape doesn't actually use all the space it asks for.
See your operating system-specific administration guide for details on adding and configuring swap space.
Question:
Running:
webmaster -w webmaster -p 9990 -u flist -f config
on jan.near.net
outputs:
Waiting for READY from 6 clients All READYs received Sending GO to all clients All clients started at Tue Aug 8 11:57:30 1995 Waiting for clients completion Reading results .Error second reading timing info from one of the clients: Interrupted system call web child 1 did not respond. 3456 bytes read .Error second reading timing info from one of the clients: Interrupted system call web child 0 did not respond. 3456 bytes read
What does the second reading timing info contain? What might cause the second read to fail while the first passes?
Answer:
It's most likely that one of the WebStone clients died before it could report results to the webmaster. We've squashed many circumstances in which this happens, but bugs continue to appear, especially on systems we haven't tested.
We can't do much for this kind of problem without debugging traces. Edit testbed, and set the DEBUG parameter to DEBUG=-d, so that debugging info will be written to files named /tmp/webstone-debug.<PID>.
If you can replicate this problem with debugging turned on, please let us know. We'd love to examine the traces.
Another possible source of problems with reading timing info is when a page in the filelist did not get read by a client, but the webmaster was expecting to find it. This can happen when the test time, number of clients and filelist distribution are set up so that a file which gets read infrequently does not get read _yet_ before the test period ends.This will get ironed out in a later release of WebStone.
WebStone primarily measures throughput (bytes/second) and latency (time to complete a request). WebStone also reports pages/minute, connection rate averages, and other numbers. Some of these may help you to sanity-check the throughput measurements.
Two types of throughput are measured: aggregate and per-client. Both are averaged over the entire test time and the entire client base. Aggregate throughput is simply total bytes (body + header) transferred throughout the test, divided by the total test time. Per-client throughput divides aggregate throughput by the number of clients.
Two types of latency are reported: connection latency and request latency. For each metric, the mean time is provided, as well as the standard deviation of all data, plus the minimum and maximum times. Connection latency reflects the time taken to establish a connection, while request latency reflects the time to complete the data transfer once the connection has been established.
User-perceived latency will include the sum of connection and request latencies, plus any network latency due to WAN connections, routers, modems, etc.
WebStone also reports a metric called Little's Ls. Ls is derived from Little's Law, and reflects how much time is spent by the server on request processing, rather than overhead and errors. Ls. is also an indirect indicator of the average number of connections which the web server has open at any particular instant. This number should stay very close to the number of clients, or else some clients are being denied access to the server at any given time.
If you load your Web servers high enough, you'll begin to see errors in the results. That's fine (at least as far as WebStone is concerned). It just means that your server is heavily loaded, and some clients aren't being serviced before they time out. In fact, the number of errors at a given load can be an excellent indicator of how your server will perform under extremely heavy loads.
Subscribe to the WebStone mailing list! Send a message to majordomo@engr.sgi.com - the subject doesn't matter, but the content should be:
subscribe webstone
You should receive a message shortly, confirming that you've been added to the mailing list. You can send to the whole list at webstone@engr.sgi.com - the authors of WebStone read the list, and they'll do their best to help. Other list members may also be able to help.
If you have access to USENET News, you can also read and post to comp.benchmarks. As with any newsgroup, read the FAQ before posting!
There's also a mailing list devoted to the performance limits of the HTTP protocol. You can subscribe by sending e-mail to www-speed-request@tipper.oit.unc.edu with the text
subscribe <your-email-address>
This file and all files contained in the WebStone distribution are copyright © 1995, 1996 Silicon Graphics, Inc.
This software is provided without support and without any obligation on the part of Silicon Graphics, Inc. to assist in its use, correction, modification or enhancement. There is no guarantee that this software will be included in future software releases, and it probably will not be included.
THIS SOFTWARE IS PROVIDED "AS IS" WITH NO WARRANTIES OF ANY KIND INCLUDING THE WARRANTIES OF DESIGN, MERCHANTIBILITY AND FITNESS FOR A PARTICULAR PURPOSE, OR ARISING FROM A COURSE OF DEALING, USAGE OR TRADE PRACTICE.
In no event will Silicon Graphics, Inc. be liable for any lost revenue or profits or other special, indirect and consequential damages, even if Silicon Graphics, Inc. has been advised of the possibility of such damages.