-none- 2007-10-02 - By Alberto Dell'Era
Back
As far as I know, Oracle doesn't communicate anything to
the OS, it is the other way around - Oracle reads OS
instrumentation data from the OS ;). Unless that has changed
in some recent version, in which case I'd like to be corrected ...
A read from a datafile (a physical read, either sequential or scattered)
is instrumented this way:
t0 := get wallclock time [ gettimeofday() system call ]
issue read request to the OS (a system call)
t1 := get wallclock time
wait time := t1 - t0;
As soon as the process issues the read system call, the OS
scheduler removes it from the CPU and puts it in the "waiting for I/O"
state. When the I/O is completed (i.e. the requested blocks have
been copied to the buffers) the process is put in the "runnable"
state (queued for the CPUs), and later put back on a CPU ("running" state).
So "wait time" includes the time spent waiting for the disks
AND the time spent in the "runnable" state (the latter is far
from being negligible if your instance has many processes
contending for the CPU) plus some time to issue gettimeofday()
etcetera.
Not sure what "CPU I/O wait" is - maybe it is simply "I/O wait",
the time a process (Oracle or not) spends waiting for the
disks, so definitely NOT on the CPU ?
Or maybe the CPU used for copying the buffers and
other activities performed inside the read system call ?
Or the two figures added together ?
BTW If you're interested in a detailed discussion, an excellent
reference is Cary Millsap/Jeff Holt's "Optimizing Oracle Performance",
where you can find these things explained in a much more
correct and precise way than mine.
HTH
Alberto
On 10/1/07, David Taft <oradbt054@(protected)> wrote:
> Qihua,
>
> Others are more knowledgeable than myself on this, but I will give this a
> shot. As I understand it, whenever Oracle is waiting on IO, the
> instrumentation is such that this time is reported to OS kernel. Hence any
> IO waits in Oracle are also reported at the OS level, but not all IO waits
> at the OS level are necessarily correlated to Oracle IO waits. When an
> Oracle process is waiting on IO to complete, it voluntarily takes itself off
> the CPU since it can't continue processing until the required IO completes.
> The Unix kernel requires this info from Oracle processes for CPU time slice
> scheduling. If I got any of this wrong, please someone correct me. Anyway,
> you may find the following article worth reading:
>
> http://www.quest-pipelines.com/newsletter-v3/0402_C.htm
> Tuning Oracle Without Cache-Hit Ratios
>
> Of particular interest with regard to your question is the section,
> IDENTIFYING THE OS BOTTLENECKS - PRONG II.
>
> You may also want to read the following post that shows a correlation
> between the OS-level IO waits and those reported in Statspack.
>
> http://www.freelists.org/archives/oracle-l/11-2005/msg00411.html
> High wio on New Hitachi SAN Storage of DB Server
>
> Hope this is of some help.
>
> Cheers,
>
> David
>
>
>
> On 9/29/07, qihua wu <staywithpin@(protected)> wrote:
> > the same or at least related to the "CPU I/O wait' at the unix level?
> >
>
>
--
Alberto Dell'Era
"the more you know, the faster you go"
--
http://www.freelists.org/webpage/oracle-l
|
|
