func runResourceTrackingTest(f *framework.Framework, podsPerNode int, nodeNames sets.String, rm *framework.ResourceMonitor,
expectedCPU map[string]map[float64]float64, expectedMemory framework.ResourceUsagePerContainer) {
numNodes := nodeNames.Len()
totalPods := podsPerNode * numNodes
By(fmt.Sprintf("Creating a RC of %d pods and wait until all pods of this RC are running", totalPods))
rcName := fmt.Sprintf("resource%d-%s", totalPods, string(uuid.NewUUID()))
// TODO: Use a more realistic workload
Expect(framework.RunRC(testutils.RCConfig{
Client: f.ClientSet,
InternalClient: f.InternalClientset,
Name: rcName,
Namespace: f.Namespace.Name,
Image: framework.GetPauseImageName(f.ClientSet),
Replicas: totalPods,
})).NotTo(HaveOccurred())
// Log once and flush the stats.
rm.LogLatest()
rm.Reset()
By("Start monitoring resource usage")
// Periodically dump the cpu summary until the deadline is met.
// Note that without calling framework.ResourceMonitor.Reset(), the stats
// would occupy increasingly more memory. This should be fine
// for the current test duration, but we should reclaim the
// entries if we plan to monitor longer (e.g., 8 hours).
deadline := time.Now().Add(monitoringTime)
for time.Now().Before(deadline) {
timeLeft := deadline.Sub(time.Now())
framework.Logf("Still running...%v left", timeLeft)
if timeLeft < reportingPeriod {
time.Sleep(timeLeft)
} else {
time.Sleep(reportingPeriod)
}
logPodsOnNodes(f.ClientSet, nodeNames.List())
}
By("Reporting overall resource usage")
logPodsOnNodes(f.ClientSet, nodeNames.List())
usageSummary, err := rm.GetLatest()
Expect(err).NotTo(HaveOccurred())
// TODO(random-liu): Remove the original log when we migrate to new perfdash
framework.Logf("%s", rm.FormatResourceUsage(usageSummary))
// Log perf result
framework.PrintPerfData(framework.ResourceUsageToPerfData(rm.GetMasterNodeLatest(usageSummary)))
verifyMemoryLimits(f.ClientSet, expectedMemory, usageSummary)
cpuSummary := rm.GetCPUSummary()
framework.Logf("%s", rm.FormatCPUSummary(cpuSummary))
// Log perf result
framework.PrintPerfData(framework.CPUUsageToPerfData(rm.GetMasterNodeCPUSummary(cpuSummary)))
verifyCPULimits(expectedCPU, cpuSummary)
By("Deleting the RC")
framework.DeleteRCAndPods(f.ClientSet, f.InternalClientset, f.Namespace.Name, rcName)
}
Client: f.ClientSet,
InternalClient: f.InternalClientset,
Name: rcName,
Namespace: f.Namespace.Name,
Image: framework.GetPauseImageName(f.ClientSet),
Replicas: totalPods,
NodeSelector: nodeLabels,
})).NotTo(HaveOccurred())
// Perform a sanity check so that we know all desired pods are
// running on the nodes according to kubelet. The timeout is set to
// only 30 seconds here because framework.RunRC already waited for all pods to
// transition to the running status.
Expect(waitTillNPodsRunningOnNodes(f.ClientSet, nodeNames, rcName, f.Namespace.Name, totalPods,
time.Second*30)).NotTo(HaveOccurred())
if resourceMonitor != nil {
resourceMonitor.LogLatest()
}
By("Deleting the RC")
framework.DeleteRCAndPods(f.ClientSet, f.InternalClientset, f.Namespace.Name, rcName)
// Check that the pods really are gone by querying /runningpods on the
// node. The /runningpods handler checks the container runtime (or its
// cache) and returns a list of running pods. Some possible causes of
// failures are:
// - kubelet deadlock
// - a bug in graceful termination (if it is enabled)
// - docker slow to delete pods (or resource problems causing slowness)
start := time.Now()
Expect(waitTillNPodsRunningOnNodes(f.ClientSet, nodeNames, rcName, f.Namespace.Name, 0,
itArg.timeout)).NotTo(HaveOccurred())
framework.Logf("Deleting %d pods on %d nodes completed in %v after the RC was deleted", totalPods, len(nodeNames),
