Allow controlling qps and burst

This adds flags for setting burst and qps. The go-code defaults
match the old zero-values which resulted in 10 and 5 being used.
Deployment default is raised to match new performance enhancements
of the gpu plugin. Values are capped to 1000 and negatives are not
accepted. Giving zeroes will result in 10 and 5 being used.

Signed-off-by: Ukri Niemimuukko <[email protected]>

Author: uniemimu

gpu-aware-scheduling/README.md

@@ -99,6 +99,8 @@ name |type | description| usage | default|
 |enableAllowlist| bool | enable POD-annotation based GPU allowlist feature | --enableAllowlist| false
 |enableDenylist| bool | enable POD-annotation based GPU denylist feature | --enableDenylist| false
 |balancedResource| string | enable named resource balancing between GPUs | --balancedResource| ""
+|burst| int | burst value to use with kube client | --burst| 10
+|qps| int | qps value to use with kube client | --qps| 5
 
 Some features are based on the labels put onto pods, for full features list see [usage doc](docs/usage.md)
 

gpu-aware-scheduling/cmd/gas-scheduler-extender/main.go

@@ -5,6 +5,7 @@ package main
 
 import (
 	"flag"
+	"fmt"
 	"os"
 
 	"github.com/intel/platform-aware-scheduling/extender"
@@ -22,13 +23,17 @@ var (
 )
 
 const (
-	l1 = klog.Level(1)
+	l1             = klog.Level(1)
+	defaultQPS     = 5
+	defaultBurst   = 10
+	maxQPSandBurst = 1000
 )
 
 func main() {
 	var (
 		kubeConfig, port, certFile, keyFile, caFile, balancedRes string
 		enableAllowlist, enableDenylist                          bool
+		burst, qps                                               uint
 	)
 
 	flag.StringVar(&kubeConfig, "kubeConfig", "/root/.kube/config", "location of kubernetes config file")
@@ -39,12 +44,22 @@ func main() {
 	flag.BoolVar(&enableAllowlist, "enableAllowlist", false, "enable allowed GPUs annotation (csv list of names)")
 	flag.BoolVar(&enableDenylist, "enableDenylist", false, "enable denied GPUs annotation (csv list of names)")
 	flag.StringVar(&balancedRes, "balancedResource", "", "enable resource balacing within a node")
+	flag.UintVar(&burst, "burst", defaultBurst, fmt.Sprintf("burst value used with kube client (limited to %d)", maxQPSandBurst))
+	flag.UintVar(&qps, "qps", defaultQPS, fmt.Sprintf("qps value used with kube client (limited to %d)", maxQPSandBurst))
 	klog.InitFlags(nil)
 	flag.Parse()
 
 	klog.V(l1).Infof("%s built on %s with go %s", version, buildDate, goVersion)
 
-	kubeClient, _, err := extender.GetKubeClient(kubeConfig)
+	for _, ptr := range []*uint{&qps, &burst} {
+		if *ptr > maxQPSandBurst {
+			klog.Warningf("Given flag value %d is too high. Limited to %d.", *ptr, maxQPSandBurst)
+			*ptr = maxQPSandBurst
+		}
+	}
+
+	kubeClient, _, err := extender.GetKubeClientExt(kubeConfig, int(burst), float32(qps))
+
 	if err != nil {
 		klog.Error("couldn't get kube client, cannot continue: ", err.Error())
 		os.Exit(1)

gpu-aware-scheduling/deploy/gas-deployment.yaml

@@ -23,6 +23,8 @@ spec:
         - "--cert=/gas/cert/tls.crt"
         - "--key=/gas/cert/tls.key"
         - "--cacert=/var/run/secrets/kubernetes.io/serviceaccount/ca.crt"
+        - "--burst=100"
+        - "--qps=50"
         - "--v=4"
         image: intel/gpu-extender
         imagePullPolicy: IfNotPresent