Re: Performance problems (solved)

Joerg Schilling <[email protected]> wrote:

> "Ruud Baart" <[email protected]> wrote:
>
> > The option fs=144m  is important for the speed of the backup. With the Sony AIT SDX-
> > 700C I managed to get a speed of +/- 20.000 kBbytes/sec.  The servers has 1 Gb of 
> > memory. The streamer has a buffer of 18 Mb.  I found hardly any speed difference 
> > between fs=90M and fs=256M. Only the fifo statistics were different. 
> >
> > Some statistics:
> >     /opt/schily/bin/star: fifo had 1836497 puts 76736 gets.
> >     /opt/schily/bin/star: fifo was 115 times empty and 572 times full.
> >     /opt/schily/bin/star: fifo held 150994944 bytes max, size was 150994944 
> >     bytes
> >     /opt/schily/bin/star: 76736 blocks + 0 bytes (total of 80463527936 bytes = 
> >     78577664.00k).
> >     /opt/schily/bin/star: Total time 8432.992sec (9317 kBytes/sec)
> >
> > We maintain 4 identical servers on 4 different locations. The overal backup speed is now 
> > between 6.000 kBytes and11.000 kBytes/sec. I did not have time to find out why 
> > 20.000 kBytes/sec is possible 
> >
> > Furthermore I found that xfsdump is the fastest dump utility. Faster then star. But very 
> > specific: only XFS filesystems.
>
> How much faster was it for you?
>
> xfsdump is a hacked ufsdump and my tests show that for an average ufs filesystem,
> star is approx. 20% faster than ufsdump.

Let me add another note:

If the OS or FS implementation for I/O through a filesystem layer is 
implemented badly, then it may be that star is slower than ufsdump or one
of it's hacked variants.

If the OS is implemented decently, then you always get more performance
from star as star benefits from the prefetch of the filesystem implementation
and the block layer of the OS.

Note that ufsdump usually will fork off 4 processes that do some kind of
prefetch and caching and as ufsdup needs to read the raw device, it does not 
see any kind of OS prefetch. Star on the other side includes a FIFO and allows
you to define any senseful size for the FIFO and thus will allow to prefetch
and cache more then 4xblocksize (what ufsdump does).

Jörg

-- 
 EMail:[email protected] (home) Jörg Schilling D-13353 Berlin
       [email protected]		(uni)  
       [email protected]	(work) Blog: http://schily.blogspot.com/
 URL:  http://cdrecord.berlios.de/old/private/ ftp://ftp.berlios.de/pub/schily