[netperf-talk] nonsense results from windows with 2.4.x / trunk

[Andrew Gallatin]

Andrew Gallatin writes:
 > 
 > I was trying to track down why netperf2.4 gives nonsense
 > results on windows when built with cygwin.  I think I have
 > tracked it down to a glibc'ism in htond()/ntohd().  
 > 
 > The problem seems to be that __BYTE_ORDER is declared on cygwin, but
 > __FLOAT_WORD_ORDER is not.  This leads to falling into the /* Fixup
 > mixed endian floating point machines */ code, which causes the
 > problem.  Plenty of other platforms (FreeBSD, Solaris, MacOSX) seem to
 > define neither, so they avoided this problem.
 > 
 > The attached patch seems to make netperf/netserver behave
 > on Windows.
 > 

Turns out that netperf is much slower (nearly a factor of 2) when
built in cygwin.  Using MS DDK way of building it makes it just as
fast (or faster) than an old netperf 2.1 binary I had laying around.

In order to get it to build, I needed to fix 2 warnings (treated
as errors) in netlib.c:access_buffer()

- dirty_totals might be used uninitialized
- adding i to *temp_buffer might result in overflow

I've attached an updated patch that gets netperf 2.4 building/running
correctly using both the cygwin and MS DDK ways of building.

Drew
-------------- next part --------------
A non-text attachment was scrubbed...
Name: winfix.diff
Type: application/octet-stream
Size: 1225 bytes
Desc: windows fixes
Url : http://www.netperf.org/pipermail/netperf-talk/attachments/20060628/b5602b94/winfix.obj