Re: [ANNOUNCE][PATCH 2.6] md: persistent (file-backed) bitmap and async writes

[Paul Clements <paul.clements@steeleye.com>]

Paul Clements wrote:
> OK, here we go again. The patch has been updated to 2.6.9-rc2 and mdadm 
> 1.7.0, and I've implemented the suggestions you've made below...
> 
> Patches: http://parisc-linux.org/~jejb/md_bitmap/

Well, I ran some simple write performance benchmarks today and I was 
stunned to find the bitmap performance was terrible. After debugging a 
while, I realized the problem was having the bitmap daemon do the 
wait_on_page_writeback...if the daemon got caught at the wrong time, it 
could take quite a while to come back around and do the writeback. So, I 
added yet another thread, the bitmap writeback daemon (whose sole 
purpose is to wait for page writebacks) and now things are running 
excellently. My simple benchmarks, which previously showed about a 
25-30% slowdown when using a bitmap vs. not using one, now show only a 
4% slowdown, which is pretty close to the margin of error in the test 
itself. Check out the new patch here:

http://parisc-linux.org/~jejb/md_bitmap/md_bitmap_2_37_2_6_9_rc2.diff

Thanks again,
Paul



> Neil Brown wrote:
> 
>> This is definitely heading in the right direction.  However I think
>> you have missed some of the subtleties of plugging.
> 
> 
> [snip]
> 
>> That should then have the bitmap updates happening before the
>> data updates, and should cluster bitmap updates better.
> 
> 
> OK, I've fixed that.
> 
>> The next step would be to stop write_page from being synchronous - I'm
>> not comfortable about ->unplug-fn having to wait for disk io.
> 
> 
> Well, it still does have to wait for the bitmap writes to complete 
> before letting the regular writes go, but I see your point about the 
> writes being submitted simultaneously (nearly) rather than being 
> completely serialized.
> 
>> This looks to be quite awkward as there is not call-back that happens
>> when writeback on a page is completed.  You need to have a kernel
>> thread to wait for the pages to complete.  I would get the
>> bitmap_daemon to do this.
> 
> 
> OK, this is what I've done...there didn't seem to be a better place to 
> do it.
> 
>> Does that make sense?
> 
> 
> Certainly does, and seems to work...
> -
> To unsubscribe from this list: send the line "unsubscribe linux-raid" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>