[PATCH v4 2/2] sched/fair: Scan cluster before scanning LLC in wake-up path

[Tim Chen]

On Fri, 2022-06-10 at 14:39 +0800, Yicong Yang wrote:
> 
> 
> I've tested this patch with MySQL as well (like in v2). This won't hurt
> the MySQL case with SIS_PROP but observed some improvement with SIS_UTIL
> posted in [1]. We leverage the nr to suppress redundant scanning in the
> current approach and seems SIS_UTIL is more efficient in this case.
> 
> 			 5.19-rc1		   patched	 patched+SIS_UTIL[1]
> TPS-16threads		  6215.11	  6172.74 (-0.68%)	  6217.33 (0.04%)
> QPS-16threads		124302.21	123454.68 (-0.68%)	124346.52 (0.04%)
> avg-lat-16threads	     2.57	     2.59 (-0.65%)	     2.57 (0.00%)
> TPS-24threads		  8726.40	  8690.87 (-0.41%)	  8833.08 (1.22%)
> QPS-24threads		174527.88	173817.42 (-0.41%)	176661.54 (1.21%)
> avg-lat-24threads	     2.75	     2.76 (-0.36%)	     2.71 (1.33%)
> TPS-32threads		  9555.42	  9514.86 (-0.42%)	 10010.87 (4.77%)
> QPS-32threads		191108.37	190297.28 (-0.42%)	200217.35 (4.55%)
> avg-lat-32threads	     3.35	     3.36 (-0.30%)	     3.20 (4.58%)
> TPS-64threads		 10290.10	 10324.75 (0.34%)	 10819.77 (5.15%)
> QPS-64threads		205802.05	206494.95 (0.34%)	216395.40 (4.90%)
> avg-lat-64threads	     6.22	     6.20 (0.38%)	     5.92 (4.88%)
> 

Thanks for the numbers. SIS_UTIL will keep off migrations off the cluster
that doesn't really improve overall utilization.  We have higher chance that
L2 cache is warm.  So it makes sense that we see a bit
better performance there.

Tim