Hi Uwe,
Do you know of any issues with XFS on Linux 7 with direct I/O?
Do you have any suggestions how to further track down the issue? E.g., how
could I prove there's something wrong with the O_DIRECT calls?
Uwe Küchler <uwe@xxxxxxxxxxxx> hat am 30. Juli 2015 um 21:14 geschrieben:--
Dear fellows of the Oracle,
From Red Hat and Oracle Linux 7 onwards, XFS is the default file system ofthe OS.
At a customer site, XFS was already the preferred file system, so the
customer chose to stick to it for a new VM with OL7 and Oracle 12.1.0.2.
But, while testing the migrated database against the old one, most of the
batch jobs showed a slowdown to at least twice the run time than in the
11.1 environment.
Both statspack reports showed clearly that "db file sequential read" was
by far the main wait event.
Top SQL and their explain plans did not differ between the environments.
Research took a while, but to get to the point: It boiled down to the I/O
response times, as shown in the wait event histogram excerpts below:
With
- 24 GiB RAM
- 5 GiB sga_target
- Buffer Cache: 4,592M
- "filesystemio_options=NONE":
Total ----------------- % of Waits
------------------
Event Waits <1ms <2ms <4ms <8ms <16ms <32ms <=1s
1s-------------------------- ----- ----- ----- ----- ----- ----- ----- -----
-----
db file scattered read 836 89.7 1.6 .8 1.8 3.1 1.9 1.1
db file sequential read 121K 83.8 .7 .8 3.9 6.1 2.8 1.9
.0
80-90% of those waits < 1ms?
This can most certainly be attributed to file system caching (no Flash
Cache, SSD or other smart stuff in place here).
- 24 GiB RAM
- 8 GiB sga_target
- Buffer Cache: 5,856M
- With "filesystemio_options=SETALL":
Total ----------------- % of Waits
------------------
Event Waits <1ms <2ms <4ms <8ms <16ms <32ms <=1s
1s-------------------------- ----- ----- ----- ----- ----- ----- ----- -----
-----
db file scattered read 63K 42.6 1.9 3.2 19.9 26.0 4.0 2.4
db file sequential read 208K 47.1 1.6 3.7 19.0 21.9 4.5 2.3
.0
In other batch job runs, the amount of waits < 1 ms was even lower (some
30%).
As you can see, I made the SGA / the buffer cache bigger in the 12c
environment, to allow for more buffering within the SGA.
Of course, I checked MOS for any known issues with Direct I/O on XFS in
this constellation, but haven't found anything so far. Just the usual
recommendations to avoid double buffering and also the confirmation that
XFS is capable of doing direct I/O.
And now for my
QUESTION(s):
============
Do you know of any issues with XFS on Linux 7 with direct I/O?
Do you have any suggestions how to further track down the issue? E.g., how
could I prove there's something wrong with the O_DIRECT calls?
Thanks for your time.
Uwe
P.S.: On (a hundred-and-) second thought I could try to enlarge the buffer
cache even more, as there's enough RAM left. At least for the tests.
P.P.S.: In case Kevin Closson reads this: I am eagerly awaiting your
upcoming blog article on XFS!
---
http://oraculix.com
