[Ocfs-users] Follow up on async I/O question

[Sunil Mushran]

The real issue is that asyncio + directio has one problem
in the 2.4 kernel, i.e., non-contiguous ios. The probable
reason it was not detected earlier was because raw, the
only core kernel component using directio, performs only
contiguous ios. This issue cannot be fixed in the 2.4 kernel
without breaking binary compatibility. It has been since
been addressed in the 2.6 kernel.

As ocfs is a filesystem, it allows files to be non-contiguous.
Also, as it is a "clustered" file system, it requires the ios
to be o_direct. So, well, we are affected by this kernel bug
or limitation, depending on whom you talk to.

So, if anyone were to ask, does ocfs support asyncio, the
easier answer is no. However, if someone persisted, the
answer is, only if your logfiles are contiguous. And as that
gets into reading debugocfs outputs, the user has to make
the determination, if the effort is worth the gain in performance.

Why only logfiles? Well, because Oracle performs large ios
only to the logfiles. The ios to be datafiles are in smaller chunks.

Hope this helps.

Sunil

Derek Suzuki wrote:

>	A few weeks back we opened a TAR with Oracle support to determine
>whether an OCFS (1.0.9-12) + async configuration was considered supportable.
>At first a technician said yes, but I followed up with Wim's explanation of
>how that combination is potentially troublesome and inadequately tested.
>The technician double-checked, then confirmed that OCFS + async is
>considered risky.  He did say that a later Oracle patch would address this.
>	Just a couple of days ago, in response to a performance-related TAR,
>another technician said that we should try turning async mode on.  When we
>referred him to the previous TAR, he said that the problems have since been
>fixed.
>	I just noticed that OCFS 1.0.11 has been released (we skipped 1.0.10
>due to the problems people were reporting).  I was wondering if this version
>explicitly addresses the known concerns about async mode, or if the
>technician was misinformed.  We do plan to test this release on a QA system,
>but I suspect that any problems would be non obvious.
>	Any further info on this matter would be appreciated.
>
>Derek
>_______________________________________________
>Ocfs-users mailing list
>Ocfs-users at oss.oracle.com
>http://oss.oracle.com/mailman/listinfo/ocfs-users
>  
>