Re: DBWR performance

From: Noons <>
Date: Sun, 13 May 2012 15:05:20 -0700 (PDT)
Message-ID: <>

On May 12, 11:51 am, wrote:

> 1. mdadm blows with asynch io+direct io: xfs+no raid does not have drop in write speed as xfs+raid1.

Slowdown in raid 1 is expected. How much slow down can be sustained then becomes the limit.

> 2. async io used in oracle is no good: it looks like it randomizes writes too much for dbwr.

Weird. It should not affect write order? I can understand the aio slaves returning control in different order but that should not affect how dbwr sequences the issuing of writes.

> 3. dbwr slaves are bad idea - interface between dbwr and slaves consumes too much cpu in dbwr.

General consensus seems to be those are useless. I guess few have ever made good use of them and why they are still available is a mistery. I suppose if one does not have any muti-path ability on the hardware/OS, then there might be a case to use them. Then again who doesn't, in this day and age?

> 4. multiple writers do not start if you have dbwr slaves specified.


> 5. multiple writers improve nothing compared to one writer for insert select and single hdd.

Expected. They are just multiplexing access to a bottleneck.

> 6. this is a winning combo so far (1 writer, no slaves, direct io):
> db_writer_processes=1
> filesystemio_options=directio
> log_buffer=81920000
> Constant write speed
> lgwr - 130MB/sec
> dbwr - 100MB/sec

Interesting the large log buffer. Any thoughts/results on/with other sizes?

> Simple is better :-), but still far from target 200.

Aye! (I know a few who disagree with the "simple" thing - but then again, who cares?)

> This is insert select then truncate with one hdd and huge db dblock buffers to make lgwr write first and then dbwr.
> I still have to see if write cache in adaptec will improve this.

You'd definitely like to use SLOB. It is ideal for this sort of test. I've used it to test/verify and fine tune memory, log writing speed and db writing speed. Have a read through what it can do and how simple it is to modify its characteristics. Received on Sun May 13 2012 - 17:05:20 CDT

Original text of this message