Oracle 8i, HP-UX 11.00, and async I/O

[VINCENT SPURGEON]

I have N-Class server running Oracle 8i and HP-UX 11. I also have two groups of Oracle dba's debating whether to implement async i/o on the server. One group is claiming that datafile corruption can occur and the other is debating that was a previous release problem.

My question, does anyone know the benefits vs. the problems with implementing async i/o? Has anyone ever done this and if so, any "gotcha's" that I should know about? Oh, and by the way, the Databases are at around 40GB in size.

[A. Clay Stephenson · Accepted solution]

By definition, data corruption can occur with asynch io. The driver is lying to the process by telling it than an I/O operation is complete
before commiting it to disk. You are trading integrity for speed. In practice, I have never been burned and the performance increase is noticeable (~20% on my boxes during heavy I/O)
but the potential for corruption is there. I would make certain that my boxes are rock solid
(power, patches, network, etc.) before using asynch I/O.
One other technique I have used is to use OnLine JFS with convosync=direct,mincache=direct mount options for datafiles and indices. This gives you all the benefits of raw files (i.e. bypassing the UNIX buffers) while still using cooked files.

[Pedro Sousa]

Hi!
I have several oracle DBs, and I use that async driver. It's better for performance.
It works with dbs as it works with NFS:
"when async is specified, the server acknowledges write requests from clients before writing data to disk. Clients do not have to wait for a write request to complete before issuing another request."

so, I would recommend you to use it, and don't forget to check the if you have the patch PHKL_18543 installed.

hope this helps.

[Pedro Sousa]

Another thing,
don't forget give the MLOCK privilege to dba group:
setprivgrp dba MLOCK

check the following thread:
http://forums.itrc.hp.com/cm/QuestionAnswer/1,1150,0x15c3d1e5762fd5118fef0090279cd0f9,00.html
good luck.

[James R. Ferguson]

Hi:

Clay's comments offer two options for you.

A good overview of the effects of the kernel's 'fs_async' parameter, including the precautions you are advised to take before setting it, can be read here:

http://docs.hp.com/hpux/onlinedocs/os/KCparam.FSasync.html

The documents "Mount options explained" (ID #KBAN00000258) and "VXFS intent log explained"
(ID #KBAN00000151) offfer good explanations of various JFS mount options, along with some suggestions for trading performance and integrity.

Together, these documents should help you make an informed choice.

...JRF...

[Alexander M. Ermes]

Hi there.
The only reason for async io is a large amount of data manipulating. Then the async io makes sense for the logfiles on raw devices.
Sample large dataloads.
Rgds
Alexander M. Ermes

[Philip P. Hartl]

From the HP-UX Performance & Tuning class; the recommendation was "mount -o fsasync" should only be used on a mount point dedicated to Oracle filespaces. The reasoning is since Oracle raw partitioning has no file system protections at all, as long as only Oracle filespaces were in the mount point, no additional risks were being taken. Once established no critical inode maintenance should be done to these filesystems, eliminating the risk of file system corruption and any data failures should be recoverable by Oracle.

If other files are allowed (add, modify, remove) the filesystem may be in motion during a crash leaving the filesystem unmountable.

[Amit Kureel]

When we are still discussing async I/O, I would like to add my queris...

I want to know how the async I/O for a raw-partition is different from file system async I/O, because normally on Unix systems it is suggested that we keep database on raw-partition to bypass unix buffering of data. But all the same Oracle ducumentation suggests us to use async I/O for its raw-partitioned database. Don't we run the risk of losing data when we have configured RAW-PARITIONs with async I/O....?

I also want to know in what cases a RAW-PARITION is better than an ONLINE JFS with WRITE-THROUGH CACHE option enabled...?

Can anybody refer me to a detailed documentation on RAW-PARTITIONING and its funcionality or some good white-papers....

Thanks in advance

Amit Kureel