fsync, ext3, raid (md) 1, write barriers and PATA caching

[Ric Wheeler]

Bruno Wolff III wrote:
> On Mon, Dec 11, 2006 at 12:44:40 -0500,
>   Ric Wheeler <ric at emc.com> wrote:
> 
>>I will send you the test code that I use & some test runs,
> 
> 
> I tried out Ric's test code and it does appear that the barrier is working
> under fc5 using software raid. For the quick test I didn't have an idle
> system. If I ran with either no barrier option and write caching enabled
> I saw very roughly a 10x speed up. The test seemed to show running with
> write caching disabled was about 20% faster than with just using barriers.
> My theory is that because there was a lot of other disk activity, the cache
> flushes forced by using barriers was writing a lot of disk blocks from
> outside the test making it report slower numbers, while in theory my system
> throughput was actually better.
> Tonight I can try the test on a system without a lot of disk activity and
> see if that makes much difference.
> 
> Thanks Ric.
> 
Glad to see that it is useful.

Thanks really go out to Jens Axboe and Chris Mason for getting this all 
to work correctly in the first place ;-)

As I mentioned in our private email exchange, you should see better 
performance with the write barrier on vs write cache disabled for some 
scenarios (large file writes and using data journal mode for small files 
are the two cases that I have noticed).

ric