On Fri, Oct 01, 2010 at 03:09:45PM -0700, Sunil Mushran
wrote:> I was discussing with Joel whether it was time for us to special
> case the odirect hack we've had since day 1. The one that allows
> concurrent odirect writes.
Right.
> I am wondering whether we should allow users control that behavior
> via a mount option. parallel_directio, maybe.
I think that's got a lot of usefulness, so yeah I'd definitely be behind
this. IMHO, we should take the opportunity to clearly define what it means
for the customer too:
"With this, parallel direct I/O from multiple nodes will be allowed. This
increases throughput on some workloads dramatically but at the expense of
inode mtime updates..." etc etc.
--Mark
--
Mark Fasheh