Oracle FAQ Your Portal to the Oracle Knowledge Grid

Home -> Community -> Mailing Lists -> Oracle-L -> RE: Miserable Disks

RE: Miserable Disks

From: Kevin Closson <>
Date: Wed, 24 May 2006 09:48:53 -0700
Message-ID: <>

>>>But even together I didn't expect things to be *this* bad.
>>>Am I looking for a non-existent fourth problem?
>>>:-) My current focus is on the 512-byte writes seen on
>>>iostat - any ideas why I could be seeing this?

No doubt you are unhappy with the performance and to your credit, Charlotte, you are not surprised by the performance. Shame on whoever chose this technology for you.

Some points. First, 2 second iostat resolution is very unreliable. Generally iostat is better on a 15-30 second interval in my experience. Second, you are complaining about 512 byte writes. I don't see those writes. What I see is 875 sectors by 35 transfers or 448000 Bytes. That is roughly an average of 8.6KB per transfer. Please tell us your Oracle block size is 8K? The .6 is likely for ext3 everhead such as metadata reads or bogus readaheads that may or may not have resulted in a bread hit. A real filesystem that supports real direct I/O (guess where you can get one of those :-) ) would not have this slight overhead. That aside, the overhead is at the noise level.

Also, (NOTE TO ALL FOLKS BECAUSE I'VE SEEN THIS "THROW ASYNC AT A DISK LATENCY ISSUE" THING ON THIS LIST A LOT), async I/O does NOT reduce disk transfer times. A DMA is a DMA, period. In fact, the
performance data you are looking at is completely unaware of whether or not the I/O request at the library level was async or not. It is just a scsi request to iostat. Async I/O only relates to the processor efficiency related to a given I/O workload.

In summary, you are not seeing 512 Byte transfers--or at least not average transfers as I'm sure there may well single sector cntrl and redo log xfers mixed in to the average.

avg-cpu: %user %nice %sys %iowait %idle

           0.64 0.00 0.13 8.71 90.52

Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s  rkB/s wkB/s avgrq-sz avgqu-sz await svctm %util

sdc          0.00  39.18 76.29 35.05 1321.65  875.26  
660.82   437.63    19.73     1.19   10.75   9.00

>>>--- Dennis Williams <>
>>>> Charlotte,
>>>> Is this directly attached to the server or is it NAS?
>>>> Dennis Williams
>>>Do You Yahoo!?
>>>Tired of spam? Yahoo! Mail has the best spam protection

Received on Wed May 24 2006 - 11:48:53 CDT

Original text of this message