[httperf] The glibc FD_SETSIZE problem and a solution

[Jesse Byler]

First of all, for httperf users who haven't found it, Julian T J 
Midgley's http performance tutorial at 
http://xenoclast.org/doc/benchmark/ is an invaluable supplement to 
the httperf documentation, as it provides explanations for how to 
interpret httperf output and how to determine if the output is valid, 
as well as a program autobench which wraps httperf to provide the 
ability to run a series of tests and graph the results.

I ran into problems when trying to implement Julian's recommendations 
at 
http://xenoclast.org/doc/benchmark/HTTP-benchmarking-HOWTO/node7.html 
for tuning file descriptor limits on Linux.  The problem is that 
increasing the file descriptor limit (either as Julian describes or 
by using the bumpfdrun wrapper at 
http://developers.sun.com/solaris/articles/polling_efficient_bumpfdrun.c ) 
doesn't buy you much with a stock httperf, because httperf-0.8 
notices that the number of file descriptors is bigger than it can 
hold in an fdset (FD_SETSIZE) and complains:

httperf: warning: open file limit > FD_SETSIZE; limiting max. # of 
open files to FD_SETSIZE

Thus a stock httperf-0.8 doesn't take advantage of the increased file 
descriptor limit.  My clunky solution was to recompile the entire 
libc with a new FD_SETSIZE and link httperf with this modified libc.  
Julian has a much simpler solution - hack httperf ("simply take a 
copy of the fd_set struct definition in sys/select.h, change the name 
of the struct, set a large FD_SETSIZE and slap it into a header file 
that you include wherever necessary in httperf's source").  He 
provided me with a patch that does exactly that.

The attached patch file (from Julian T J Midgley, submitted here by 
permission - thanks Julian!) should apply against a clean httperf-0.8 
tree:

 cd httperf-0.8; patch -p1 < /tmp/fix_fd_setsize.patch

Optionally set LOCAL_FD_SETSIZE in the new file fd_set.h as required 
(it defaults to 102400, which is likely to be high enough for most 
purposes), rebuild, and the patched httperf does not suffer from the 
glibc FD_SETSIZE limitation.

While we're in the business of patching the httperf-0.8 source, I'll 
include another minor patch which fixes the ./configure problem where 
httperf likes to install into /bin instead of $prefix/bin.  If there 
are any httperf maintainers on this list, perhaps the two patches 
could be incorporated into an httperf-0.9 release?

Thanks,
Jesse
-------------- next part --------------
A non-text attachment was scrubbed...
Name: fix_exec_prefix.patch
Type: text/x-diff
Size: 338 bytes
Desc: not available
Url : http://napali.hpl.hp.com/pipermail/httperf/attachments/20051128/93ca3f44/fix_exec_prefix.bin
-------------- next part --------------
A non-text attachment was scrubbed...
Name: fix_fd_setsize.patch
Type: text/x-diff
Size: 3208 bytes
Desc: not available
Url : http://napali.hpl.hp.com/pipermail/httperf/attachments/20051128/93ca3f44/fix_fd_setsize.bin