Quoting Justus Winter (2014-12-19 11:19:34) > Quoting Samuel Thibault (2014-12-19 02:54:17) > > Justus Winter, le Fri 19 Dec 2014 02:48:56 +0100, a écrit : > > > Use the ternary operator to implement `assert' like it is done in the > > > glibc. The glibcs changelog does not mention the rationale behind > > > this change, but it does improve our IPC performance. > > > > Mmm, is it not a matter of changing the estimation of probability of > > branching from the compiler? I guess using an "unlikely" macro would > > get the same benefit in a much more readable and compiler-portable way? > > Yes, I tried that first, without getting any improvements. The libc > once used __builtin_expect for assert too, but stopped (they didn't > mention why in the changelog). > > Lacking a specialized IPC microbenchmark I'm using iperf to measure > TCP performance (over netdde), which is greatly influenced by IPC > performance. > > I'll work a little on my IPC microbenchmark, and try to reproduce my > findings with it.
I have pushed my benchmark here: http://darnassus.sceen.net/gitweb/teythoon/machometer.git/ Here is an example run: start ext2fs: Hurd server bootstrap: ext2fs[device:hd0s1] exec startup proc auth. + dhclient -v /dev/eth0 Internet Systems Consortium DHCP Client 4.3.1 Copyright 2004-2014 Internet Systems Consortium. All rights reserved. For info, please visit https://www.isc.org/software/dhcp/ can't create /var/lib/dhcp/dhclient.leases: Read-only file system Listening on Socket//dev/eth0 Sending on Socket//dev/eth0 DHCPDISCOVER on /dev/eth0 to 255.255.255.255 port 67 interval 7 DHCPREQUEST on /dev/eth0 to 255.255.255.255 port 67 DHCPOFFER from 10.0.2.2 DHCPACK from 10.0.2.2 /sbin/dhclient-script: 52: /sbin/dhclient-script: cannot create /etc/resolv.conf.dhclient-new: Read-only file system chown: cannot access '/etc/resolv.conf.dhclient-new': No such file or directory chmod: cannot access '/etc/resolv.conf.dhclient-new': No such file or directory mv: cannot stat '/etc/resolv.conf.dhclient-new': No such file or directory can't create /var/lib/dhcp/dhclient.leases: Read-only file system bound to 10.0.2.15 -- renewal in 37346 seconds. + iperf -t 60 -c 10.0.2.2 ------------------------------------------------------------ Client connecting to 10.0.2.2, TCP port 5001 TCP window size: 64.0 KByte (default) ------------------------------------------------------------ [ 3] local 10.0.2.15 port 32768 connected with 10.0.2.2 port 5001 [ ID] Interval Transfer Bandwidth [ 3] 0.0-60.0 sec 855 MBytes 119 Mbits/sec + /machometer -n 33554432 N: 33554432 (1<<25), qlimit: 5 mach_print: 5s840000us 174.045563ns 5745621.918 (1/s) nullmsg: 6s400000us 190.734863ns 5242880.000 (1/s) producer: 28s420000us 846.982002ns 1180662.632 (1/s) consumer: 28s420000us 846.982002ns 1180662.632 (1/s) + /machometer -n 33554432 N: 33554432 (1<<25), qlimit: 5 mach_print: 5s820000us 173.449516ns 5765366.323 (1/s) nullmsg: 6s390000us 190.436840ns 5251084.820 (1/s) producer: 23s570000us 702.440739ns 1423607.637 (1/s) consumer: 23s570000us 702.440739ns 1423607.637 (1/s) + /machometer -n 33554432 N: 33554432 (1<<25), qlimit: 5 mach_print: 5s810000us 173.151493ns 5775289.501 (1/s) nullmsg: 6s420000us 191.330910ns 5226547.040 (1/s) producer: 25s670000us 765.025616ns 1307145.773 (1/s) consumer: 25s670000us 765.025616ns 1307145.773 (1/s) Too bad our clock is soo bad :/ Also, I'm fed up with cpusets to pin my qemu process. Measuring the impact of the patches will have to wait. Justus
