[Csync2] linux kernel cpu optimization

[Vincent Régnard]

Giampaolo Tomassoni a écrit :
>> -----Original Message-----
>> From: csync2-bounces at lists.linbit.com [mailto:csync2-
>> bounces at lists.linbit.com] On Behalf Of Vincent Régnard
>> Sent: Thursday, September 11, 2008 3:03 PM
>> To: csync2 at lists.linbit.org
>> Subject: [Csync2] linux kernel cpu optimization
>>
>> Hi all,
>>
>> I have performance issues using csync2 on SMP system
>> I csync2 a big disk whith ~ 1 million files, no frequent updates, only
>> small changes on ext3 fs + linux 2.6.16 kernel. Cyncing is very long
>> (somtimes half an hour), hardware is efficient, process should
>> certainly
>> run better than that.
>> I forced all disk interrupt on CPU1 using the
>> /proc/irq/IRQ/smp_affinity
>> trick, but I observe that CPU1 remains in waiting state too much as
>> shown by the top above (about 50% on average) when csync is checking
>> the
>> whole file system tree.
>>
>> I execute csync2 with a nice priority (level 5)
>> top - 12:54:08 up 6 days,  1:33,  4 users,  load average: 0.80, 0.65,
>> 0.66
>> Tasks: 139 total,   4 running, 135 sleeping,   0 stopped,   0 zombie
>> Cpu0  :  0.0% us,  9.4% sy,  2.3% ni, 66.9% id, 21.4% wa,  0.0% hi,
>> 0.0% si
>> Cpu1  : 11.3% us, 31.7% sy,  6.3% ni,  0.0% id, 49.7% wa,  0.3% hi,
>> 0.7% si
>> Mem:    903100k total,   893576k used,     9524k free,    68436k
>> buffers
>> Swap:  1959920k total,        8k used,  1959912k free,   312908k cached
>>
>> How can I improve that situation ? Can I tweek kernel params to change
>> that or is it csync2 internals specific ? NB: I do not have similar
>> behaviour with other heavy processes (such as zipping/unzipping huge
>> files, backup process etc..).
> 
> If you're experiencing a low CPU usage but disk access is fast, problems may
> be around the network. What ifconfig is saying about your net interface? In
> particular, have a look to the errors/dropped/overruns/frames counters in
> the interface you use to csync. Do it on both nodes.
> 
> Giampaolo
> 

I think network is not a real problem in my case, or at least is not 
responsible for this observed cpu latency. Without any network traffic 
between hosts (controlled with tcpdump), I observe waiting ratios over 
70% like this:

Cpu1  :  0.3% us, 17.4% sy,  3.0% ni,  0.0% id, 76.6% wa,  1.3% hi,  1.3% si
10800 root      22   5  4600 3564  768 R 45.6  0.4   0:20.14 csync2

Result is similar with and without nicing the process.

The large values of waiting ratios are observed when csync2 is accessing 
disk.

-- 
Vincent Régnard
vregnard at tbs-internet.com
TBS-internet.com
027 630 5902

-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/x-pkcs7-signature
Size: 5793 bytes
Desc: S/MIME Cryptographic Signature
Url : http://lists.linbit.com/pipermail/csync2/attachments/20080911/6c52378a/smime-0001.bin