Re: [PATCH] eliminate large inline's in skbuff

On Wednesday 28 April 2004 00:21, Stephen Hemminger wrote:

quoted

This takes the suggestion and makes all the locked skb_ stuff, not inline.
It showed a 3% performance improvement when doing single TCP stream over 1G
Ethernet between SMP machines. Test was average of 10 iterations of
iperf for 30 seconds and SUT was 4 way Xeon.  Http performance difference
was not noticeable (less than the std. deviation of the test runs).

The original tests were suspect for a whole lot of reasons.  Running the proper
tests shows no performance differences.  The best theory as to why there was
a difference in earlier tests is that memory debugging was enabled;  that caused
each buffer to be overwritten with a memset. When that happened, the test
ends up measuring the speed of the memory and cache bandwidth, not the CPU or
the network.

On Thursday 06 May 2004 00:59, Stephen Hemminger wrote:

quoted

On Wednesday 28 April 2004 00:21, Stephen Hemminger wrote:

quoted

This takes the suggestion and makes all the locked skb_ stuff, not
inline. It showed a 3% performance improvement when doing single TCP
stream over 1G Ethernet between SMP machines. Test was average of 10
iterations of iperf for 30 seconds and SUT was 4 way Xeon.  Http
performance difference was not noticeable (less than the std. deviation
of the test runs).

The original tests were suspect for a whole lot of reasons.  Running the
proper tests shows no performance differences.  The best theory as to why
there was a difference in earlier tests is that memory debugging was
enabled;  that caused each buffer to be overwritten with a memset. When
that happened, the test ends up measuring the speed of the memory and cache
bandwidth, not the CPU or the network.

"no detectable difference" was the most expecter result indeed.
Thank you for actually testing that.
--
vda