Re: Log file I/O throughput

From: Yong Huang <yong321_at_yahoo.com>
Date: 13 Aug 2003 18:27:18 -0700
Message-ID: <b3cb12d6.0308131727.7bbd7913@posting.google.com>

yong321_at_yahoo.com (Yong Huang) wrote in message news:<b3cb12d6.0308130744.6f16d704_at_posting.google.com>...
> Sybrand Bakker <gooiditweg_at_sybrandb.demon.nl> wrote in message news:<tbfjjvk38fc03uilc960baks99b7gh08jg_at_4ax.com>...
> > >
> > >I know some people don't like it. But using one log member per group
> > >when you have good hardware redundancy is actually good practice. It
> > >saves on log file single write time (when AIO is enabled). I remember
> > >Steve Adams recommended it (or could be somebody else?) and also
> > >noticed a person working at BMC on the OAUG mailing list was doing
> > >that. I've been doing that for some time. You just have to be a little
> > >more careful in deleting files.
> > >
> > >Yong Huang
> >
> >
> > Sorry to contradict you. When of the members of your hardware mirror
> > fails, the write request to the redolog file fails, and your database
> > will crash.
> > Also, when you have multiple cpus, and you did multiplex your
> > redologs, the archiver will use both to assemble the archived redolog.
> >
> >
> > Sybrand Bakker, Senior Oracle DBA
>
> Hi, Sybrand,
>
> Are you saying when the RAID system fails to write to one mirror
> member, the other mirror member won't be written to either?
>
> The second part. Are you saying if you have 2 members per group, the
> archiver reads both members simultaneously and somehow assembles the
> content just read into one archive log file? Anyway, hardware
> mirroring already gives you read performance boost.
>
> I found Steve Adams' notes at
> http://www.ixora.com.au/tips/use_raw_log_files.htm. He says "Hardware
> mirroring can be used in preference to redo log multiplexing to avoid
> serial writes to multiple log file members, and indeed such hardware
> mirroring is recommended to minimize the CPU cost of redo writes." I
> think in the first part he was referring to serial writes to _headers_
> of log file members, because writing to log file bodies would be
> parallelized. But parallel writes still cost CPU (his second part).
>
> Yong Huang

Important addition. With Steve Adams' approval, I'm posting his emails here (all text included):

---

His first mail: No, the first part means "if async I/O is not available, another way of getting parallel writes is to use hardware mirroring".

Then my email: ... Most DBAs I know still configure multiple log members even if
they have hardware mirroring. I'm curious to know whether that truely helps
protect data or it's more out of paranoid.

His second mail: You may post a follow-up if you wish. I'm too busy to buy into the thread myself. I do, of course, agree with you.

---

I still hope Steve can have some time and post an in-depth analysis to this thread. In the meantime, I apologize for misreading his article.

Yong Huang Received on Wed Aug 13 2003 - 20:27:18 CDT