> Have you done any profiling of real workloads to see where the actual
> overhead comes from?
Not yet. I've spent the last weeks learning enough about the code to
make sense of profiles :)
This week (probably wednesday) I'll put both my netfilter hook statistic
patch, and enabled kernel profiling, onto a production box (the transproxy
thing from the bucket occupation analysis). Right now I have totally
undersized bucket count on that machine (7168 buckets for 10 times
the tuples), so I'll first measure the "accidental long list walk"
situation, and then retry with a suitable bucket size.
best regards
Patrick
|
