Skip site navigation (1)Skip section navigation (2)
Date:      Sat, 19 Oct 96 14:42:29 -0500
From:      Ben Black <black@gage.com>
To:        "Matthew N. Dodd" <winter@jurai.net>
Cc:        fs@freebsd.org
Subject:   Re: ccd setup for striping 
Message-ID:  <9610191942.AA01614@squid.gage.com>
In-Reply-To: <Pine.BSI.3.95.961018204227.13041B-100000@sasami.jurai.net>
References:  <Pine.BSI.3.95.961018204227.13041B-100000@sasami.jurai.net>

next in thread | previous in thread | raw e-mail | index | archive | help
--NeXT-Mail-466589079-1
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Content-Disposition: inline

>When you have 150,000 customers news is a high availability application.
>(I don't have this problem yet, but know people who do.)

when you have 150k customers you have more than one news server.

>I'd like to see your hard numbers as to the 'software RAID killing
>performance'.  Its had the opposite effect here.  Of course, we're not
>running RAID5, just interleaving.

"just interleaving" doesn't give you any redundancy, now does it?  RAID 5 is  
all about parity calculations.  we used sun online disk suite to do a  
software RAID on one of several news servers.  performance was far better  
simply distributing the news groups across drives intelligently.

>I'm sorry, but a high end news server is seek bound.
>
>Run some tests on a test platform before you make statements about
>performance.

yes, assuming you aren't recalculating parity every time to read or write  
data, it is seek bound.  to eliminate the massive seek bottleneck, get rid of  
the filesystem.  that's what we did.  as i've said before, 6 million  
articles per day on a P5-120 with 64MB RAM.  then we ran out of articles.   
load average under 2.0.  i know about new server performance.  any questions?


b3n

--NeXT-Mail-466589079-1
Content-Type: text/enriched; charset=us-ascii
Content-Transfer-Encoding: 7bit
Content-Disposition: inline

<smaller>>When you have 150,000 customers news is a high availability
application.
<nofill>
>(I don't have this problem yet, but know people who do.)

when you have 150k customers you have more than one news server.

>I'd like to see your hard numbers as to the 'software RAID killing
>performance'.  Its had the opposite effect here.  Of course, we're not
>running RAID5, just interleaving.
</nofill>

"just interleaving" doesn't give you any redundancy, now does it?  RAID 5
is all about parity calculations.  we used sun online disk suite to do a
software RAID on one of several news servers.  performance was far better
simply distributing the news groups across drives intelligently.

<nofill>
>I'm sorry, but a high end news server is seek bound.
>
>Run some tests on a test platform before you make statements about
>performance.
</nofill>

yes, assuming you aren't recalculating parity every time to read or write
data, it is seek bound.  to eliminate the massive seek bottleneck, get rid
of the filesystem.  that's what we did.  as i've said before, 6 million
articles per day on a P5-120 with 64MB RAM.  then we ran out of articles. 
load average under 2.0.  i know about new server performance.  any
questions?



b3n

</smaller>

--NeXT-Mail-466589079-1--



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?9610191942.AA01614>