Re: Tuning ZFS: performance anomaly with nocacheflush

From: Charles Schultz <sacrophyte_at_gmail.com>
Date: Wed, 19 Nov 2008 11:45:27 -0600
Message-ID: <7b8774110811190945x49706354ga61c01c9f49e41f8@mail.gmail.com>


Thanks. "Unfortunately" we are running u5.

On Wed, Nov 19, 2008 at 10:57 AM, David Miller <David.J.Miller_at_sun.com>wrote:

> Hi Charles,
>
> I asked one of the zfs experts and his comment was:
>
> nocacheflush shouldn't be a problem since S10u5. If they are on
> S10u4 or u3 (ack!) then they are probably hitting one or more of
> other known problems, long since fixed.
>
> So check what version of Solaris you're running and you may need to
> upgrade.
>
> Regards,
>
> Dave Miller
>
> Charles Schultz wrote, On 11/18/08 09:44:
>
>> Good day list,
>>
>> We are investigating why turning on the ZFS nocacheflush would affect a
>> simple (traditional) import so drastically. With the nocacheflush default
>> (0), we observe normal behavior. With the nocachflush set to 1, we observe
>> each db writer process suddenly having a LWP (Lightweight Process) count in
>> excess of 150 (some up to 175). This caused significant kernel contention
>> and slowed down the import significantly. A truss on dbwr processes shows
>> lots of sleeping and lwp_parking. I was not able to find much relevant
>> information on google or metalink for tags "oracle zfs nocacheflush". Any
>> ideas?
>>
>> Oracle EE 10.2.0.2 <http://10.2.0.2> on Solaris 10, hardware is Sun
>> T5240, EMC C3-80 SAN.
>>
>> --
>> Charles Schultz
>>
>

-- 
Charles Schultz

--
http://www.freelists.org/webpage/oracle-l
Received on Wed Nov 19 2008 - 11:45:27 CST

Original text of this message