On Sun, May 4, 2014 at 3:26 PM, Wojciech Puchar <[email protected]> wrote: >> For what it's worth, Sepherosa Ziehau has been putting a lot of work into >> reducing CPU use on high-bandwidth connections: >> http://lists.dragonflybsd.org/pipermail/users/2014-March/128240.html >> > if i understand correctly it would reduce CPU usage greatly on highly loaded > server doing lots of TCP transfers.
There is another commit to reduce port token contention. Recent TCP connect(2)+kqueue(2) using 8 processes could do ~300Kconns/s; compared w/ ~275Kconns/s before the port token splitting work. The TCP connect(2)+kqueue(2) is cpu bound on Dfly now (~0% idle time on all HTs). Admittedly, I still can't max out 1000M network w/ connect(2) on my testing hardware (the theoretical TCP connect rate for 1000M network should be ~335Kconns/s). However, on the other hand, the TCP accept(2)+kqueue(2) is quite good w/ SO_REUSEPORT, sink @~335Kconns/s w/ pretty much idle time on each HT. Best Regards, sephe > > looks like core i7 with 4 cores and 8 threads can saturate completely 2 > 10Gbit/s ethernet interfaces with both directions. > > better than FreeBSD, but anyway - it's still over 3 CPU cycles per byte > transmitted. still very high CPU usage > > > >> http://lists.dragonflybsd.org/pipermail/commits/2014-March/269653.html >> >> >> On Sat, May 3, 2014 at 6:54 AM, Wojciech Puchar >> <[email protected]> wrote: >> as my friend advised me, i did flood test - generated lots of random >> UDP packets with size between 52 and 100 bytes. >> >> Actually low end ethernet card in my laptop was a limit (rl), but >> compared FreeBSD 10 with custom kernel and dragonfly 3.6 with generic kernel >> (that have diagnostics check >> compiled by default). >> >> dragonfly used about 3 times less CPU. >> >> no tests on gigabit or 10 gigabit ethernet and multicore machine for >> now. >> >> quite good. >> >> CPU load on more normal cases (like transfering files using rcp) is >> lower too, but difference is smaller like 1.5-2 times. >> >> >> >> > -- Tomorrow Will Never Die
