Asynchronous I/O for uncooked file system - Informix

[Michael Callaghan]

I have read the forum about Asynchronous I/O and it is mainly directed to oracle cooked filesystems. The question I have is, has anyone had experience with Asynchronous I/O and uncooked filesystems for Informix. The feedback I have received so far is that it is not advantagous on informix and that it was more trouble than it was worth. If anyone has experience in this particular scenario, whether you thought it was of major benefit or whether the advantages by doing this were minimal. Thanks

[Steve Lewis]

I assume that by aync i/o you really mean KAIO and by uncooked filesystem you mean raw to LVM or a filesystem which bypasses buffers.

Nearly all informix reads are asynchronous.
The only synchronous writes are checkpoints (unavoidable) and foreground writes (undesirable). Apart from that, the database either uses its own AIO virtual processors to service requests or KAIO if you configure it.

KAIO puts a reduced load on the server due to fewer processes, but increases the system load in kernel routines.

The issue with AIO VPs is that they all get scheduled with every other process on the server (more waits), whereas KAIO goes straight to the kernel from the CPU VP that requests it.

I have found that KAIO improves performance, but doesn't work well on more than one instance per machine. Keep the less important instances as non-KAIO. Also you should export IFMX_HPKAIO_NUM_REQ=3000 or so for a busy instance.

If you want cooked files, it is important that you bypass the HP-UX buffer cache, otherwise the logical and physical log records might not get flushed to disk if the server TOCs or PANICs, so you would end up losing integrity.

On old operating systems (<11.xx) and older versions of IDS (<7.3) the KAIO wasn't so good but these days its fine.

[Michael Callaghan]

Thanks I am going to reward points but I wanted to see if I could get more feedback

I do mean KAIO and raw filesystems for informix

Does anyone else have any feedback ??

[Doug_3]

I use KIAO and found the perf to be beneficial. On the iiug.org listserv there are a number of posts, all of which are favorable.

The issue I think is relevant includes the type of disk sub-system you are using. Is it a set of JBOD's, an xp512, something with lot's of memory cache, multiple pathing, etc. Raw disk io does eliminate the OS buffering and overhead. But if you are using 10k disks on a very intelligent san or disk sub-system or multiple dbspaces, fragmentation and pdq, the impact maybe neglible. Your milage may vary!!!!!

Also, if you use cooked files, there are those who say you can use a OS back up utility (Veritas, tar, fbackup, etc) with success; giving you a secondary means of back up and disaster recovery. I have no exp with this but am just passing along 2nd hand info.

Hope this helps,
Doug

[Michael Callaghan]

Thanks for that

We are using an EMC symm
So that is one for one against

Is there any other imput

both of your imputs have been helpful

[Michael Callaghan]

Is there any other input because I have one for and one against