Overview speeds and methods

Christian Balzer chibi at gol.com
Tue Aug 13 07:09:30 UTC 2002

Kondou-san wrote:
>In article <E17eQzW-0007fQ-00 at batzmaru.gol.ad.jp>,
>	Christian Balzer <chibi at gol.com> wrote;
>} 1. Which is considered the overall fastest overview method? (I read that 
>} classic is best for readers)
>It depends on your enviroment.  How many articles
>are you fed?  

About 850000 at 100GB per day.

>How many simultaneous readers do you
Not many, about 50 at complete peak, average is more like 20.

>I've noted many times on the list that the
>reason why I wrote buffindexed is that tradindexed
>was too slow to handle fullfeed on my test
>environment(signle disk :).  ovdb is said to be
>fast but I've never compared to both on the same
>environment.  I was quite happy with my own
>implementation at my previous job, while others
>often reported on the list that buffer is corrupted.
Speaking of which, from todays expire run:
Aug 12 05:35:39 nnrp expireover[27337]: buffindexed: ovgroupmmap ovbuff is 
null(ovindex is 25938, ovblock is 154154291
Aug 12 05:35:39 nnrp expireover[27337]: buffindexed: could not open overview 
for 'alt.stupid.morons'

>} 2. Will buffindexed as a regular file benefit from caching as opposed to
>} the buffers being raw devices?
>I haven't tried buffers on raw devices, but it may
>have.  I had 4GB memory and 20*2GB buffers for
>overview(5% used) in the box, and over 400 simultaneous
>readers, it's fed up to 300GB/day.  It worked without
>any particular stress.
Sounds impressive. I guess if that error up there doesn't go away
I will rebuild the overview with tradindexed or ovdb and see how
that fares.


Christian Balzer        Network Engineer        Engineering
chibi at gol.com   	Global OnLine Japan/Exodus Communications K.K.

More information about the inn-workers mailing list