Re: [coLinux-devel] autobuild org.colinux.devel 20080907 12:26:17

[Henry N. <hen...@ar...>]

Hello Harald,

Harald Weidner wrote:
>> http://www.henrynestler.com/colinux/autobuild/devel-20080907/
>>      
> [...]
>    
>> Add sysenter / sysexit restoration support
>>      
> [...]
>    
>> If host CPU supports sysenter (SEP), then coLinux kernel will use
>> sysenter now. Sysenter is faster as the old Int 0x80.
>>      
>
> I updated from Build 20080823 to 20080907 and expected coLinux to
> become faster. However, the opposite seems true.
>
> My standard "benchmark" is a compile run of the plain Linux kernel
> 2.6.18, doing "make clean; make defconfig; time make all". I issued
> three runs each. Here are the results (real/user/sys min:sec):
>
> coLinux Snapshot 20080823
> 1. 6:17 / 4:25 / 1:51
> 2. 6:15 / 4:28 / 1:49
> 3. 6:14 / 4:26 / 1:45
>
> coLinux Snapshot 20080907
> 1. 6:30 / 4:26 / 2:03
> 2. 6:32 / 4:26 / 2:04
> 3. 6:30 / 4:28 / 2:01
>
> Building the Linux kernel is mainly a CPU and memory intensive task;
> however, there are also many small files created and removed, which
> invokes syscalls.
>
> The host machine has an Intel Core2 Duo 6400 processor at 2.13 GHz
> and 2 GB of RAM, running an (otherwise unloaded) Windows XP Professional
> SP 3. coLinux has 512 MB of memory and runs Debian 4.0r4 with the
> kernel shipped with coLinux.
>
>    
I have seen also such effects. After coLinux boots up: 3x build OpenVPN, 
every with unpacking source, configure and make.

Version 0.8.0-20080823:
51.911s / 26.770s / 12.160s
29.330s / 16.470s / 11.380s
29.290s / 15.040s / 12.890s

Version 0.8.0-20080907:
37.811s / 14.900s / 10.600s  (Faster from Windows file cache after 
coLinux reboot)
29.460s / 16.240s / 12.030s  (+0.4%)
29.580s / 16.520s / 11.740s  (+1%)
( real / user / sys )

In vanilla Linux sysenter should run faster as Int80. But, nobody says 
how mutch faster.

CoLinux needs to save and restore 3 more special registers now. This are 
3*RDMSR + 3*WRMSR on every operating system switch, typically hardware 
interrupts for harddisk and timer, and for memory allocation or free. 
This cost also some time. I don't know how many. The counts of switching 
can read from /proc/colinx/stats (stats_clear clears the counters after 
read). I assume, that the time we got from sysenter, will lose on OS 
switch now.

Currently I'm not sure, is the distribution using sysenter? I only know, 
that the sysenter exist from kernel. I have tested with [1]. I have not 
idea how can check, that glibc or the elf-loader use the sysenter. For 
example a staticly linked old program will use Int 0x80.

Someone found, that syscalls on Pentium 4 are slower as Pentium III [2,3].
In [4] is a good summary about sysenter inside Linux kernel.

[1] http://www.trilithium.com/johan/2005/08/linux-gate/
[2] http://kerneltrap.org/node/531
[3] http://marc.info/?l=linux-kernel&m=103942289025149&w=2
[4] http://manugarg.googlepages.com/systemcallinlinux2_6.html

-- 
Henry N.