On Wed, 2016-10-26 at 20:09 +0200, Thomas Gleixner wrote:

On Wed, 26 Oct 2016, Tim Chen wrote:

quoted

On Wed, 2016-10-26 at 13:24 +0200, Thomas Gleixner wrote:

quoted

quoted

There were reservations on the multi-socket case of ITMT, maybe it would
help to spell those out in great detail here. That is, have the comment
explain the policy instead of simply stating what the code does (which
is always bad comment policy, you can read the code just fine).

What is the objection for multi sockets? If it improves the behaviour then
why would this be a bad thing for multi sockets?

For multi-socket (server system), it is much more likely that they will
have multiple cpus in a socket busy and not run in turbo mode. So the extra
work in migrating the workload to the one with extra headroom will
not make use of those headroom in that scenario.  I will update the comment
to reflect this policy.

So on a single socket server system the extra work does not matter, right?
Don't tell me that single socket server systems are irrelevant. Intel is
actively promoting single socket CPUs, like XEON D, for high densitiy
servers...

Instead of handwaving arguments I prefer a proper analysis of what the
overhead is and why it is not a good thing for loaded servers in general.

Then instead of slapping half baken heuristics into the code, we should sit
down and think a bit harder about it.

The ITMT scheduling overhead should be small.  Mostly a small number of 
cycles initially spent to idle balance tasks towards an idled favored core, and cycles to refill
hot data in the mid level cache for the migrated task.  Those should be a very
small percentage of the cycles that the task spent running on the favored core.
So any extra boost in frequency should compensate so should be a good trade off.

After some internal discussions, we think we should enable the ITMT feature by 
default for all systems supporting ITMT.  I will remove the single socket
restriction.

Thanks.

Tim