Database Reference
In-Depth Information
[...]
lat (usec) : 500=0.01%, 750=0.01%, 1000=0.01%
lat (msec) : 2=99.84%, 4=0.13%
The number of IOPS has increased from 18,628 to 24,595, and the execution time is down to 5,329 milliseconds
instead of 7,036, but this has come at a cost. Instead of microseconds, 99.84% of the I/O requests now complete in
milliseconds.
When you are working with storage and Linux, FIO is a great tool to help you understand the storage solution
as well as the hardware connected against it. FIO has a lot more to offer with regards to workloads. You can read
and write sequentially or randomly, and you can mix these as well—even within a single benchmark run! The
permutations of I/O type, block sizes, direct, and buffered I/O, combined with the options to run multiple jobs
and use differently sized memory pages, etc., make FIO one of the best benchmark tools available. There is a great
README as part of FIO, and its author has a HOWTO document, as well, for those who want to know all the detail:
Oracle I/O numbers
The ORION tool has been available for quite some time, initially as a separate download from Oracle's website and now
as part of the database installation. This, in a way, is a shame since a lot of its attractiveness initially came from the fact
that a host did
not
require a database installation. If you can live with a different, lower, version number, then you can
still download ORION for almost all platforms from Oracle's OTN website. At the time of this writing, the below URL
The above link also allows you to download the ORION user's guide, which is more comprehensive than this section.
For 11.2 onwards you find the ORION documentation as part of the official documentation set in the “Performance
Tuning Guide”. Instead of the 11.1 binaries downloadable as a standalone package, this topic will use the ones provided
with the Oracle server installation.
ORION works best during the evaluation phase of a new platform, without user-data on your logical
unit numbers (LUNs). (As with all I/O benchmarking tools, write testing is
destructive
!) The package will use
asynchronous I/O and large pages, where possible, for reads and writes to concurrently submit I/O requests to
the operating system. Asynchronous I/O in this context means the use of
libaio
on Linux or equivalent library as
discussed in the FIO benchmark section above. You can quite easily see that the ORION slave processes use
io_submit
and
io_getevents
by executing
strace
on them:
[root@server1 ~]# strace -cp 28168
Process 28168 attached - interrupt to quit
^CProcess 28168 detached
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ----------------
97.79 0.323254 4 86925 io_submit
2.11 0.006968 1 12586 io_getevents
0.10 0.000345 0 25172 times
------ ----------- ----------- --------- --------- ----------------
100.00 0.330567 124683 total
[root@server1 ~]#
The process with ID 28168, to which the strace session attached, was a running ORION session.
Search WWH ::
Custom Search