Re: OT: high WIO on Linux

Kevin,

I've always wondered that. How can you determine that? Is the iostat average IO size per device any good for this ?

Sometimes even if it does chew it up in 64k requests, I find it's better to have it large, so that you can queue up a large number of IOs in a serial read process.

On 3/15/06, Kevin Closson <kevinc_at_polyserve.com> wrote:
>
> Are the depths of my I/O queues in any way "tunable" under
> Linux? (I'm pretty sure I recall tuning this under AIX...) I would
> expect that Linux will make some rather bland assumptions about the SCSI
> or FC "disks" it finds in my (hypothetical) RAID-array. Or is it smart
> enough to query the device in the hope of learning the best queue depth
> to use... (I have no doubt that the SCSI standards permit such queries,
> but that does not necessarily mean that the OS will bother to ask, nor
> that the device will provide a "truthful" answer.) Anyway, I would
> expect a RAID-array with a few GB of NVRAM to be able to support queue
> depths into the tens of thousands. Is that irrational?
>
> ...this topic always turns into a bit of a cognitive reaction test
> really. Here's the
> deal. We are talking about Oracle generating this I/O demand, not a
> microbenchmark. So,
> if you have blisteringly fast I/O (e.g., SSD, or an oversized SAN array
> cache), you will
> be able to satifsy the I/O requests in sub millisecond service times.
> When I/O is really
> fast for Oracle, it goes CPU bound. The only way to get Oracle to issue
> more I/O in that
> case is to either get a faster server, or shave off some of the code
> that Oracle is
> executing around the I/Os. To summarize, a really fast I/O subsystem
> makes queue depth
> a bit of a non-issue.
>
> What about the other way around? That is, a slow I/O subsystem. What
> effect does that
> have on queue depth? If your downwind I/O subsystem can't drain the
> I/Os, it doesn't
> really help much to queue up.
>
> Queue-depth is tunable on Linux, but there is a balance that must be
> struck between
> queueing and servicing. The bigger worry you should have is whether
> your Linux
> distro is taking Oracle's multiblock IOs (direct path read/write,
> scattered read,
> CCF,etc) and chopping them up into a bunch of little (e.g., 32KB, 64KB)
> chunks
> in the scsi mid-layer.
>
>
> --
> http://www.freelists.org/webpage/oracle-l
>
>
>

--
Christo Kutrovsky
Senior Database/System Administrator
The Pythian Group
--
http://www.freelists.org/webpage/oracle-l