Re: ORACLE on Linux - IO bottleneck

Wyvern wrote:
>
> # iostat -x 2 /dev/sdc1
>
>
> Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s rkB/s wkB/s
> avgrq-sz avgqu-sz await svctm %util
> sdc1 6833,34 114,06 7902,86 95,05 14743,20 209,10 7371,60
> 104,55 1,87 34,12 4,26 0,12 93,15
>
>
> cpu-med: %user %nice %sys %iowait %idle
> 15,23 0,00 2,11 11,32 71,34
>
>
> Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s rkB/s wkB/s
> avgrq-sz avgqu-sz await svctm %util
> sdc1 7149,99 25,01 8151,98 35,52 15295,47 134,07 7647,73
> 67,03 1,88 32,95 4,03 0,11 92,75
> ---------------------------------------------------------------------------­-----------------------------------
>

 From what I read here: your system is asking for 512byte (wsec/wkB and rsec/rkB) read and writes.
Physically the writes are done in 2k (w/wkB) while the read in less of 1k (7647,73/8151,98).
The service time is low (0,11) probably due to the really short reads but it result in an inefficient operation.

If the I/O is sequential you would benefit of an I/O scheduler...

> I´ve only shown /dev/sdc statistics because almost ALL the RAWs are in
>
> that device.
>
> Well, now the questions (¡¡of course!! ;-) ):
>
> 1.- There is a generic low performance at the system, at SO level,
> at oracle level and at application level. ¿Can we suppose that there
> is an IO bottleneck?

Possible but not sure.
You are using your disks poorly, sorry, :(

> 2.- ¿Are the "wrqm/s" / "rrqm/s" values correct?
>

No, they are low. You are not merging the I/O properly (maybe because your reads are not sequential).

>
> I've been reading different manuals and documentation about Raw Devices
>
> and Direct I/O in Linux (from Oracle an Redhat). Everything makes me
> think that oracle
> blocksize (8K) should not affect performance drastically because we use
> DirectIO.
>

Direct I/O is not always the solution...

> Well, we´ve been analizing diferent statistics from Storage system
> (ECM - Symmetrix) and the number of IOs the /dev/sdc device is doing is
>
> near hardware limit (about 8000) and the medium IO size of all of them
> is
> more or less 2k. I don´t know why this happen when all the IO against
> this
> device is done by oracle and oracle have an 8K db_block_size.
>

The storage statistics are right for writes. Check if the reads are done in about 1k.

> ¿Any idea?
>

Just to be sure: check the readahead of your device.

( The /sys pseudo filesystem contains the information. Even hdparm can give you the number).

-- 
Fabrizio Magni

fabrizio.magni_at_mycontinent.com

replace mycontinent with europe