VXFS mount options with Oracle

[Eric Guerizec]

Hello,

I want to use the following vxfs mount options with the Oracle datafiles :
convosync=direct,mincache=direct,nodatainlog,delaylog,rw
I read on this forum that it's the best solution with Oracle datas filesystems.

Oracle is not up yet on my server but if I make a test with command cp, the copy takes more time with these options!!!

Examples :
to copy a 96 Mbytes file, 9 seconds but only 3 seconds with standard vxfs mount options.

to copy a 285 Mbytes file, 34 seconds but only 20 seconds with standard vxfs mount options.
I don't understand the results!
Is it normal ??????????

[Pete Randall]

Eric,

The difference is that Oracle uses raw volumes and your test is using cooked volumes. The options recommended for Oracle are to defeat the cacheing and logging that is done for "regular" file systems since Oracle handles that.

That's a pretty simple take on it but I hope it helps.

Pete

[Jon Finley]

As I see it, you'e geting rid of the file system buffering and allowing Oracle to write directly to disk (as much as it can anyway).

Oracle will take care of the internal buffering of writes. The problem with the file system buffering the writes is that once Oracle releases control of the write to the OS, it believes the transaction to be complete. But.. the file system may NOT have done the write yet.

There is a possibility of losing the transaction with file system buffering.

Jon

[A. Clay Stephenson]

That is exactly what you would expect. Because cp is using no buffer cache, it takes longer. What you are trying to avoid is the OS using buffers and Oracle using buffers in the SGA.

You only want to use the =direct options for datafiles and indices; redo and archive logs should still use the buffer cache.

Now, if you are running HP-UX 11.11, you will be better advised to use cooked files for everything; on 10.20 and 11.0 bypassing the buffers is usually the better performer. The only way you will know for sure is to measure for yourseld using Oracle. The good news is that if you put the datafiles and indices in one filesystem and the redo/archive logs in another; you can easily measure performance by simply changing the mount options.

[Eric Guerizec]

Thanks everybody!!!

[Bill Hassell]

As mentioned, these options remove the double buffering that would occur when Oracle *and* HP-UX buffer reads and writes. Read-ahead (which is what HP-UX does for buffered reads) is probably going to read a lot of unseable data from the Oracle datafiles because HP-UX has no way to know if the reads are sequential. Oracle does know what it needs to buffer and how much, so removing the opsystem buffer will improve the Oracle I/O speeds.

HOWEVER: do NOT use the convosync,mincache and nodatainlog options for your separate archive, rollback and redo log directories (and these must be separate, ideally on physically separate disk channels. You need the opsystem buffering (just like the cp command--which is sequential) for these logs as they are very busy--and sequential.