Re: V$FILESTAT Average Write Times

"Ethan Post" <Blah_at_Blah.com> wrote in message news:<VX2q7.16536$6c5.483277_at_news1.rdc1.sdca.home.com>...
> I have a number of databases I am responsible and I am starting to look at
> average file write times in V$FILESTAT. Some of the systems are using EMC
> and some are using IBM's Shark disk storage system. I am seeing average
> write times in the .5 to 1 second range on most of the systems. This seems
> high but I am dealing with an ERP application that heavily indexes tables.
> The systems are RAID 0+1 as far as I know. Read times are usually .002-.009
> seconds. What type of average write times are you used to seeing and would
> this concern you. No one is complaining and I have not noticed any real
> slowdowns. One of the systems with an exceptionally high average write time
> is experiencing a redo log bottleneck. Would this increase the average
> write time in v$filestat? Do the disk systems write cache skew these
> statistics?
>
> Thanks,
> Ethan Post

I will leave most of your questions to others but I want to make or ask about two points:

The last I knew the IBM Shark is a Raid 5 box and that is all it is. IBM did not have stripping or simple mirrowing at the hardware level so are you doing OS level stripping and mirrowing using a RAID 5 device? That is exactly what we had to do though IBM was working on the other features. Performance was actually quite good under this arrangement.

The average IO time of v$filestat can be distorted by a bad IO so the numbers are not necessarily reliable. You need to compare the Oracle stats to device stats obtained at the OS level and eliminate any periods where hung IO's and other problems existed as opposed to just load related problems.

And yes the disk unit signalling the OS that the write is done when it is only cached does affect the statistics; however, I would not consider this skewing because as far as the OS and Oracle are concerned the IO is complete. If the system is write intensive enough to flood the write cache so that the disk unit (EMC box or Shark) has to wait on the physical IO to actually be written then this will result in longer IO's and the averages will reflect it.

• Mark D Powell --