[svlug] Sync command slowdown

[James Sparenberg]

On Monday 24 November 2008 18:17:15 Luke S Crawford wrote:
> James Sparenberg <james at linuxrebel.org> writes:
> >     I've got 4 new DB servers exhibiting a rather problematic set of 
> > actions with disk sync.  Running the sync command (as is done when 
logs 
> > are rotated.)  takes 4-6 seconds (not micro, full seconds)  Other 
than 
> > this these boxes are running 10x faster than the servers they 
replaced 
> > and making our customers much happier.
> 
> Are you using ext3? (or something else that does write-back caching?)
> I saw something very similar the other night when I was setting up my 
new
> ridiculous[1] servers (I switched to 32GB ram/dual socket quad-core 
systems
> for my Xen hosting setup.)  -  See, normally I limit the Dom0 to 
having 
> 512M or a gigabyte of ram, but these were new servers and I was just 
playing
> with them in the garage.  I hadn't even installed the xen kernel yet, 
so 
> the OS instance had a whole 32GB ram to play with.  I ran some disk 
tests, 
> a dd command, bonnie, -  the system screamed.  Sync, on the other 
hand, 
> took north of 120 seconds to return.  
> 
> but on another console, iostat showed that my sata disks were still 
chugging
> away, even though the disk tests finished some time ago.  As far as I 
can 
> tell, it was just the expected ext3 write caching, just, you know, 
bigger.   
> it takes a long time to flush 32GB of data to SATA disks.  
> 
> But yeah, that'd be my guess;  you just have lots of dirty buffers.    
More ram
> means that you can buffer more, and disks are still pretty slow.  
> 
> 
> [1]http://prgmr.com/~lsc/luke_opterons.jpg
> 

Yes I am using ext3 and in the process I also found this at kerneltrap.

http://bugzilla.kernel.org/show_bug.cgi?id=9546

James