On 11/1/2010 8:35 PM, Michael Ellerman wrote:

On Wed, 2010-10-27 at 12:54 -0400, Chris Metcalf wrote:

quoted

The tile architecture uses this framework for our serial console,
and our users complain that the delay of up to two seconds feels like
the machine has gone non-responsive and is disturbing.  By contrast,
a delay of up to half a second feels like just the normal sort of
delay caused by swapping, network lag, etc. and is not noticeable.
The overhead is obviously not much greater.

It's 4 times greater.

We picked 2 seconds because it gave a reasonable trade off between
responsiveness and load. I'm not convinced that half a second is a
better number.

Perhaps the tradeoff should be tunable, then?  I think on our architecture
we're willing to pay a higher cost on the core running this task, since we
have many cores; we often have a core that mostly just runs miscellaneous
Linux administrative tasks anyway, so adding a bit more overhead there is
not significant for us.

This issue has caused multiple reports of unresponsiveness from our users,
so I'd be interested in finding a way to strike a balance.  We could use a
config option defaulting to 2 seconds (for example), or something more
dynamic (probably unnecessary).

Let me know your preference, if this sounds plausible, and I'll write up a
proposed patch.  Thanks.

-- 
Chris Metcalf, Tilera Corp.
http://www.tilera.com