Skip site navigation (1)Skip section navigation (2)
Date:      Tue, 14 Jan 97 16:56:51 CST
From:      Joe Greco <jgreco@solaria.sol.net>
To:        henrich@crh.cl.msu.edu (Charles Henrich)
Cc:        dg@root.com, terry@lambert.org, freebsd-hackers@FreeBSD.ORG
Subject:   Re: mount -o async on a news servre
Message-ID:  <199701142256.QAA03967@solaria.sol.net>
In-Reply-To: <199701122054.PAA14676@crh.cl.msu.edu> from "Charles Henrich" at Jan 12, 97 03:54:19 pm

next in thread | previous in thread | raw e-mail | index | archive | help
> Hmm.
> 
> Let me say what I feel is how a generaly large all-purpose news spool does per
> day.
> 
> Create's and writes about 200k files and virtually zero directories.
> Reads about 1-2 million files and sends them off to the net.
> And delete's about 200k files a day (during one, hopefully,  very small time
>     period)
> 
> A large number of reads should be satisfied out of memory cache (presuming you
> fill the box with enough memory).
> 
> Setting noatime means that the duration of the sync() call in update() is
> markedly less (as obviously there is 1/5th to 1/10th the work to do).
> 
> Now if I can push back the update() where it actually is pushed back (how?)
> there should be a fairly large percentage of updates occuring to the same
> directories, which should give me less update I/O's over the same time period.
> Meaning that when reads come in that must hit a disk, the chances of hitting a
> bottleneck because a update() is hitting all disks in the farm, is 10% as
> likely to happen.

Wow.  :-)  Ditto what he said.

I guess I don't have enough intimate knowledge of ffs to say how it works,
but based on observation, on a busy machine:

If I set update to 30, right after a sync event, I get an average of about
15 articles/sec processed, and that rises to about 25 over the next 30
seconds, and then BOOM a sync happens, and we drop down to 15 again.

If I set update to 300, right after a sync event, I again do about 15 a.p.s.,
and over the period of the next minute, ramp up to about 45 articles per
second...  which I sustain for the NEXT FOUR MINUTES (i.e. next sync).

If the "ramp up" is linear, that means that I process about 6000 articles in
300 seconds if update is 30, or 12600 articles if update is 300.

It's sort of like freeway driving vs. driving on the side streets.  Less
stopping means you get there faster.

... JG



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199701142256.QAA03967>