On Sat, 2016-01-30 at 15:20 +0100, Frederic Weisbecker wrote:
> On Fri, Jan 29, 2016 at 05:43:28PM -0500, Rik van Riel wrote:
> > Run times for the microbenchmark:
> >
> > 4.4 3.8 seconds
> > 4.5-rc1 3.7 seconds
> > 4.5-rc1 + first patch 3.3 seconds
> > 4.5-rc1 + both patches 2.3 seconds
>
> Very nice improvement!
Tasty indeed.
When nohz_full CPUs are not isolated, ie are being used as generic
CPUs, get_nohz_timer_target() is a problem with things like tbench.
tbench 8 with Rik's patches applied:
nohz_full=empty
Throughput 3204.69 MB/sec 1.000
nohz_full=1-3,5-7
Throughput 1354.99 MB/sec .422 1.000
nohz_full=1-3,5-7 + club below
Throughput 2762.22 MB/sec .861 2.038
With Rik's patches and a club, tbench becomes nearly acceptable.
---
include/linux/tick.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
--- a/include/linux/tick.h
+++ b/include/linux/tick.h
@@ -184,7 +184,7 @@ static inline const struct cpumask *hous
static inline bool is_housekeeping_cpu(int cpu)
{
#ifdef CONFIG_NO_HZ_FULL
- if (tick_nohz_full_enabled())
+ if (tick_nohz_full_enabled() && runqueue_is_isolated(cpu))
return cpumask_test_cpu(cpu, housekeeping_mask);
#endif
return true;
