On Fri, 07 Nov 2014 01:19:45 +0900 (JST) YASUOKA Masahiko <yasu...@openbsd.org> wrote: > On Tue, 4 Nov 2014 14:50:49 +1000 > David Gwynne <da...@gwynne.id.au> wrote: >> this uses siphash to protect the in_pcb hashes. this is pretty much >> a textbook example of what siphash should be used for. >> >> tests? ok? > > Giving inpcb_hash() to LIST_* macros via INPCB*HASH() caused calling > inpcb_hash() multiple times unintentionally. Fixed. Diff is as > follows. > > Otherwise ok yasuoka. test yasuoka.
I put kernel gprof results on http://yasuoka.net/~yasuoka/oldhash-gprof.txt.gz http://yasuoka.net/~yasuoka/siphash-gprof.txt.gz tested on - load: 4,000 HTTP connections / sec by IxLoad - 1 HTTP transaction / connection - HTTP response body size = 1 byte - relayd(8) relays them with below config http protocol "http" { tcp backlog 32 } relay "HTTP_relay" { listen on 127.0.0.1 port 80 protocol http forward to destination } - pf.conf pass in quick inet proto tcp to port 80 divert-to localhost port 80 - profiling is enabled 10 seconds - OpenBSD 5.6 (amd64) + diff - Intel Xeon E3-1220 v3 @ 3.10GHz (4 CPUs) short result compare to pf_test(), it was 57% % egrep '^\[(62|45)]' siphash-gprof.txt [45] 1.0 0.03 0.37 682175 pf_test [45] [62] 0.5 0.00 0.21 501212 in_pcbhash [62] %
