Patrick Ohly
33ea278c51
DRA: use v1beta1 API
...
No code is left which depends on the v1alpha3, except of course the code
implementing that version.
2024-11-06 13:03:19 +01:00
Patrick Ohly
0130ebba1d
DRA scheduler: refactor "allocated devices" lookup
...
The logic for skipping "admin access" was repeated in three different places. A
single foreachAllocatedDevices with a callback puts it into one function.
2024-11-01 13:23:28 +01:00
Patrick Ohly
bd7ff9c4c7
DRA scheduler: update some log strings
2024-11-01 13:23:11 +01:00
Patrick Ohly
bc55e82621
DRA scheduler: maintain a set of allocated device IDs
...
Reacting to events from the informer cache (indirectly, through the assume
cache) is more efficient than repeatedly listing it's content and then
converting to IDs with unique strings.
goos: linux
goarch: amd64
pkg: k8s.io/kubernetes/test/integration/scheduler_perf
cpu: Intel(R) Core(TM) i9-7980XE CPU @ 2.60GHz
│ before │ after │
│ SchedulingThroughput/Average │ SchedulingThroughput/Average vs base │
PerfScheduling/SchedulingWithResourceClaimTemplateStructured/5000pods_500nodes-36 54.70 ± 6% 76.81 ± 6% +40.42% (p=0.002 n=6)
PerfScheduling/SteadyStateClusterResourceClaimTemplateStructured/empty_100nodes-36 106.4 ± 4% 105.6 ± 2% ~ (p=0.413 n=6)
PerfScheduling/SteadyStateClusterResourceClaimTemplateStructured/empty_500nodes-36 120.0 ± 4% 118.9 ± 7% ~ (p=0.117 n=6)
PerfScheduling/SteadyStateClusterResourceClaimTemplateStructured/half_100nodes-36 112.5 ± 4% 105.9 ± 4% -5.87% (p=0.002 n=6)
PerfScheduling/SteadyStateClusterResourceClaimTemplateStructured/half_500nodes-36 87.13 ± 4% 123.55 ± 4% +41.80% (p=0.002 n=6)
PerfScheduling/SteadyStateClusterResourceClaimTemplateStructured/full_100nodes-36 113.4 ± 2% 103.3 ± 2% -8.95% (p=0.002 n=6)
PerfScheduling/SteadyStateClusterResourceClaimTemplateStructured/full_500nodes-36 65.55 ± 3% 121.30 ± 3% +85.05% (p=0.002 n=6)
geomean 90.81 106.8 +17.57%
2024-11-01 13:23:06 +01:00
