Re: [JBD] change batching logic to improve O_SYNC performance

[Ric Wheeler <ric@emc.com>]

Benjamin LaHaise wrote:
> Hello folks,
> 
> When writing files out using O_SYNC, jbd's 1 jiffy delay results in a 
> significant drop in throughput as the disk sits idle.  The patch below 
> results in a 4-5x performance improvement (from 6.5MB/s to ~24-30MB/s on 
> my IDE test box) when writing out files using O_SYNC.  Instead of always 
> delaying for 1 jiffy when trying to batch, merely do a yield() to allow 
> other processes to execute and potentially batch requests.  Additionally, 
> limit the delay to an absolute max of 1/50th of a second.  
> 
> 		-ben
> 

I was playing with similar code in reiserfs last year, trying to tune 
for fsync() performance.

The challenge is that when the rate of fsync() arrivals is low, you 
don't want to introduce extra latency for these "lonely" requests while 
you wait for a request to combine which is not going to come any time soon.

On the other hand, if the rate of fsync() requests is high enough, you 
gain a huge performance win by combining multiple fsync()'s into one 
transaction.

This feels a bit like the IO combining logic in the io scheduler layer.

It would seem to be begging for some dynamic logic based on the arrival 
rate for both the fsync & synchronous write combining...

ric