To Directio or Note

[Steve Slade]

I have just found out about a Solaris option, called forcedirectio. This parameter is stored in the fstab and essentially causes the system to write directly to the disk and avoid the cache. This brings with it performance enhancements for Sequential Write Activity, (only).

Looking at the HP documentation I see mention of this for a specific RAID controller, but in general not much. It seems that the HP version of this command is directio and it something which is configured at the Array controller stage.?.

Why I am interested? We currently have an N-class machine with an AutoRaid, (HPC 5447A I believe), and I wish to know if it is possible to do something like this with our architecture. This could very useful for our database, (redo and arch), mountpoints as they are used only for sequential writes.

Any advice gladly taken.


Thanks

Steve

[S.K. Chan]

The vxfs filesystem does have such similar features that you wanted. It the matter of trying these features and options out in your mount table. Personally I have not fully explore all of them. Take a look at ..
# man mount_vxfs
# man vxtunefs
The latter is able to run on-the-fly on a mounted FS which is a neat tool. You really would want to look at the various mount options and then use vxtunefs to adjust the performance. You might want "nolog" but it's risky as log replay is not possible.

[S.K. Chan]

This explains the various mount options.. see JRF attachment of doc #KBAN00000258

[S.K. Chan]

This explains the various mount options.. see JRF attachment of doc #KBAN00000258
http://forums.itrc.hp.com/cm/QuestionAnswer/1,,0x91f6a12d6d27d5118fef0090279cd0f9,00.html

[A. Clay Stephenson · Accepted solution]

Actually I run those AutoRAID's with Oracle now. The vxfs_options that you are looking for are convosync=direct,mincache=direct,nodatainlog,delaylog.

The =direct bypass the buffer cache. Now, you actually want the redo and archive logs to use the buffer cache but the datafiles and indices to bypass the buffer cache.

If you are using 11.11 (rather than 11.0) then I find the better performance is to use the buffer cache for everything but under 11.0 (or 10.20) archive/redo logs use the buffer cache and datafiles/indices should not. If you bypass the buffer cache you can also reduce the size of the buffer cache and use that memory to expand the SGA - where Oracle does its buffering.

Because, you are probably running 11.11, use the buffer cache and bump it up to about 800-1000MB (about 400 MB max under 11.0).

Now the real trick to getting your AutoRAID's to perform is to dive the I/O. Despite the myths that you may hear, there are no advantage to going beyond 2 LUN's per VG - I've measured it. Each VG should be comprised of 2 equally sized LUN's. LUN0's primary path should be thru controller X (alternate Y) and LUN1's primary path should be thru controler Y (alternate X). Each LVOL should then stripe across both LUN's (usually in 64K chunks especially for vxfs regardless of block size).
This will allow you to fully utilize the two external SCSI busses and the 12H's four internal SCSI buses. The other key to performance is to allocate no more than 50% or so of the space as LUN's. This will keep the AutoRAID in RAID 1/0 mode at all times. If you follow these steps, you will be quite pleasantly surprised at how well these old girls perform.

[Dietmar Konermann]

Steve,
see the man page of mount_vxfs(1M). The "mincache=direct" and the "convosync=direct" options could be of interest for you. These options are available with the OnlineJFS product.

Best regards...
Dietmar.

[Steve Slade]

Many thanks to everyone for your gems of info.

This certainly looks quite interesting, and maybe worth a bit more investigation by us.

Thanks again

Steve