multiblock vs direct path read on 10gR2 aix (related: iostat/awr stat diff)

From: Jeremy Schneider <jeremy.schneider_at_ardentperf.com>
Date: Thu, 4 Aug 2011 09:43:51 -0500
Message-ID: <CA+fnDAZvyKPcTj8F8ij7cSGwV+WZrwYkfdjNH9hCMZN5g9v+9A_at_mail.gmail.com>

Interesting. I have a different question related to the same issue of I/O on 10gR2 AIX. I wonder if anyone else on this platform can tell me whether you've seen anything similar.

I have observed about a 10x difference in throughput for direct-path reads vs multi-block reads of the same size. In other words, as soon as a query switches from multiblock to direct path, it suddenly does I/O literally 10 times faster. I expect direct path to be faster (bypassing buffer cache overhead and all), but this seems a bit extreme! I'm curious about others' experiences with direct path reads - has anyone else seen this big of a speed difference between direct path and multiblock?

At one point, full tablescans were pulling table data at about 16MB/s with direct path and about 180MB/s with direct path. I haven't started digging yet, but does anyone know if these two methods use the same system I/O call? If so, that might indicate something in the database accounting for the difference, otherwise it could be a combination of DB and OS. At the storage tier, these tests were reading the exact same blocks. I alternated between them several times to make sure I wasn't just seeing cache interference.

Also, I was calculating these throughput number from the wait events... and I didn't see any wait events related to contention. All the wait time was "db scattered read" or "direct path read" - both reading 32 x 16k blocks per wait.

-Jeremy

On Wed, Aug 3, 2011 at 9:59 AM, joshuasingham <joshuasingham_at_gmail.com>wrote:

> Hi ,
>
> I been looking at some (10.2.0.4) on AIX 64 bit Awr report File IO Stats
> and can see that the Av Rd(ms) for some of the files are consistently above
> 50 to 60 ms over a 15min interval but when I reported this storage guy he
> mention and sent me the Iostat to prove that the service time in the same
> timeline is average of 20ms for the disk involve . What can cause the
> difference between the AWR and iostat readings
>
> thanks
>

-- 
http://www.ardentperf.com
+1 312-725-9249

Jeremy Schneider
Chicago

--
http://www.freelists.org/webpage/oracle-l

Received on Thu Aug 04 2011 - 09:43:51 CDT