[OLK-6.6] Make Cluster Scheduling Configurable · Pull Request !5854 · openEuler/kernel

Cluster scheduling domain was introduced in 5.16 to help even out load
between the clusters. In a last level cache, there can be multiple 
clusters, with each cluster having its own resources and multiple CPUs
in it. With cluster scheduling, contention on cluster resource (e.g. L2
cache) can be reduced for better performance.

These patches made cluster scheduling configurable at run time and
boot time.  When system is moderately loaded, it is worthwhile to do the
extra load balancing to balance out load between the clusters to reduce
contention on cluster resources (e.g. L2 cache).  If the system is
fully utilized, load balancing among cluster is unlikely going to help
to reduce contention of resources a cluster as the cluster
is fully busy.

On a Jacobsville system with 24 Atom cores, where 4 Atom core per cluster
share an L2, we ran the mcf benchmark from very low load of 1 benchmark
copy to 24 benchmark copies on the 24 CPUs system.  We see that
throughput is boosted for medium load but there is little improvement
from cluster scheduling when the system is fully loaded.

Improvement over baseline kernel for mcf_r
     copies         run time        base rate
     1              -0.1%           -0.2%
     6              25.1%           25.1%
     12             18.8%           19.0%
     24             0.3%            0.3%

If the system is expected to operate close to full utilization, the sys
admin could choose to turn off the cluster feature to reduce scheduler
overhead from load balancing at the cluster level.

Cluster scheduling is disabled by default for x86 hybrid CPUs in the
last patch of this series. For such asymmetric system, the system
should rely strictly on CPU priority to determine the order
of task scheduling.

https://gitee.com/openeuler/kernel/issues/I9F5WO

GVPopenEuler/kernel关闭

搜索帮助

GVP openEuler/kernel
关闭