RE: Large discrepancy between 'log file parallel write' and 'db file parallel write' times
Date: Thu, 3 Jan 2013 09:42:03 +0000
Message-ID: <DUB114-W8852C875D54906F552EEABB7210_at_phx.gbl>
thanks John,
in this case it turned out that we were taking advantage of Smart Flash Logging. I hadn't realised it was enabled in this environment. so the 'log file parallel writes' were hitting the flash disks not the hard disks. Matt
From: john.clarke_at_centroid.com
To: mccmx_at_hotmail.com; oracle-l_at_freelists.org
Subject: Re: Large discrepancy between 'log file parallel write' and 'db file parallel write' times
Date: Mon, 31 Dec 2012 02:01:28 +0000
Redo log and control file writes are considered "high priority" I/O consumers and get lumped into the high priority background I/O queue, so cellsrv will satisfy these requests at a higher priority than anything else.
DBWR writes (which your db file parallel write waits is associated with) are prioritized the same as user I/O, relatively speaking, and a handful of other background-related writes and lower than log file/control file writes.
Under heavy I/O load and saturation, this behavior should be what you may expect.
- John
From: Matt McClernon <mccmx_at_hotmail.com>
Reply-To: "mccmx_at_hotmail.com" <mccmx_at_hotmail.com>
Date: Sunday, December 30, 2012 5:51 PM
To: "oracle-l_at_freelists.org" <oracle-l_at_freelists.org>
Subject: Large discrepancy between 'log file parallel write' and 'db file parallel write' times
Oracle 11.2.0.2 EE on Exadata X2-2 quarter rack with extra cells (7 cells in total).
Under heavy I/O load (i.e. when the cells are saturated) we are seeing a very large difference between the average latencies of 'log file parallel write' and 'db file parallel write' waits.
From a 30 minute AWR report in which each cell was performing its theoretical maximum 1500 IOPS we observe the following write latencies:
DBFPW = 263ms
LFPW = 2ms
when the I/O subsystem is not pushed to its limits the values for both wait events are in the order of single digit milliseconds.
So my question is why do the DBWR write times suffer so much during I/O bottlenecks when the LGWR write times do not..?
The LGWR writes will be significantly smaller and are sequential as opposed to the DBWR writes which are random writes, but are those differences enough to justify the large gap, or is it possible that we are losing 'db file parallel write' time to some other processing..?
and are the LGWR writes genuinely sequential..? because ultimately the redo logs are on the same physical disks as the datafiles because both the RECO and DATA diskgroups are carved from the same cell disks.
Matt
--
http://www.freelists.org/webpage/oracle-l
-- http://www.freelists.org/webpage/oracle-lReceived on Thu Jan 03 2013 - 10:42:03 CET