SO_BUSY_POLL for latency critical cases

[Francis Daly]

On Fri, Feb 03, 2017 at 12:47:45PM +0200, Donatas Abraitis wrote:

Hi there,

> Sorry, what is `with.min` and `without.min` in this case?

Without trying to speak for Maxim:

The github link provided says "Benchmarks example. Port 6379 is without
busy polling, 6380 with:" and provides some lists of "min" and "max"
values.

I imagine that with.min is "the list of min values, with busy polling";
and the other filenames correspond to the other three sets.

And the "ministat" analysis suggests that the "with" and "without"
sets of values are statistically identical.

So the evidence provided suggests that there is no benefit from adding
SO_BUSY_POLL. "changes are really reasonably visible" is incorrect.


There may well be benefits from adding SO_BUSY_POLL. But the list
of numbers provided does not show those benefits, according to this
specific analysis.

Maybe it can be explained why the analysis is wrong; maybe some other
analysis can be suggested that will show that there are benefits. But
the justification for the change will probably have to come from the
proposer.

Cheers,

	f
-- 
Francis Daly        francis at daoine.org