Skip to main content
Version: Next

Evaluate YuniKorn Performance with KWOK

KWOK is a powerful toolkit designed to swiftly establish a cluster of thousands of Nodes within seconds. This page is dedicated to leveraging KWOK for performance testing. The objective is to conduct a comparative analysis between YuniKorn (Standard mode/Plugin mode) and Kube-scheduler, evaluating their capabilities in workload handling.

KWOK conserves memory by emulating only node and pod behavior, unlike Kubemark, which emulates a kubelet and consumes a significant amount of memory.

Using Kwok, we replicated previous performance tests, deploying 10 deployments on 5000 nodes with 5000 replicas each, resulting in a remarkably low total memory usage of 30-40GB. As a result, Kwok enables us to scale up experiments and assess the performance of YuniKorn without the need to consider bandwidth and Kubelet processing speed.

Environment

The test is conducted using KWOK in a Cluster. The cluster environment is optimized according to the performance tuning settings in the YuniKorn documentation. For more details, refer to the Benchmarking Tutorial. You can conveniently set up Kwok in your Kubernetes Cluster by downloading the scripts we provide here.

For data monitoring, Prometheus will be employed to gather metrics. We'll use count(kube_pod_status_scheduled_time{namespace="default"}) as an indicator of throughput.

Test Cases

We will start a comparative analysis with kube-scheduler to evaluate the throughput of these two different schedulers. In addition, we will compare the performance differences when managing large numbers of Taints and Tolerations, configuring Affinity and Anti-Affinity settings, and handling PriorityClass jobs.

important

YuniKorn schedules pods according to the application. Bearing this in mind, our testing focuses on assessing the impact of various features like Taint & Tolerations, Affinity, and PriorityClass on performance. This requires assigning unique configurations to each pod to understand their effects accurately. We use shell scripts to configure a large number of pods efficiently, storing their configurations in a YAML file. These pods are then deployed as needed to evaluate the impact on system performance.

During this process, we've identified that the primary constraint is the processing capability of the api-server. Our data indicates that it is capable of deploying 5,000 pods in just 38 seconds, achieving a throughput of 131.6 pods per second.

Throughtput

In this experiment, we will use the following three test cases to measure the throughput and scheduling duration of different schedulers. Each application will be deployed at one-second intervals.

Test Case:

Test CaseApplicationsTasksTotal Pods
#1150005000
#2510005000
#3252005000

Result:

#1kube-scheduleryunikornyunikorn plugin mode
makespan99699
throughput(pods/sec)50.5833.350.5
throughput-result-01
#2kube-scheduleryunikornyunikorn plugin mode
makespan999100
throughput(pods/sec)50.5555.650
throughput-result-02
#3kube-scheduleryunikornyunikorn plugin mode
makespan9932100
throughput(pods/sec)50.5156.350
note

Regarding the throughput of the YuniKorn Scheduler, the third test took the longest time. This can be attributed to the fact that the application is deployed only every second, thereby lengthening the makespan.

throughput-result-03

Taint & Tolerations

In the Taint & Tolerations test case, each node is initially assigned a taint that corresponds to its numerical identifier. Following this, pods are randomly assigned tolerations based on the indexes of different nodes. To assess the impact on system performance, we will deploy a YAML file containing 5000 pods and monitor the outcome.

kubectl taint nodes kwok-node-1 key-1=value-1:NoSchedule
tolerations:
- key: key-$rand
operator: "Exists"

Result:

Taint & Tolerationskube-scheduleryunikornyunikorn plugin mode
makespan993699
throughput(pods/sec)50.5138.950.5
taint-tolerations-result

Affinity & Non-Affinity

In the test cases for Affinity & Non-Affinity, 5,000 pods will be divided into four different combinations as follows:

Types of Node affinity and anti-affinityOperatorNumbers of Pods
PreferredIn625
PreferredNotIn625
RequiredIn625
RequiredNotIn625
Types of Pod affinity and anti-affinityOperatorNumbers of Pods
PreferredIn1250
PreferredNotIn1250

In the Node affinity section, the matchExpressions value is set to the hostname of a randomly selected node.

affinity:
nodeAffinity:
requiredDuringSchedulingIgnoredDuringExecution:
nodeSelectorTerms:
- matchExpressions:
- key: kubernetes.io/hostname
operator: $operator
values:
- kwok-node-$randHost

In the Pod affinity section, the value of matchExpressions is a randomly assigned applicationId.

affinity:
podAffinity:
preferredDuringSchedulingIgnoredDuringExecution:
- weight: 100
podAffinityTerm:
labelSelector:
matchExpressions:
- key: applicationId
operator: $operator
values:
- nginx-$randAppID1
topologyKey: kubernetes.io/role

Result

Affinity & Non-Affinitykube-scheduleryunikornyunikorn plugin mode
makespan9942100
throughput(pods/sec)50.511950
taint-affinity-result

PriorityClass

Firstly, deploy 100 distinct PriorityClasses. Subsequently, deploy Jobs with 5,000 Pods at the same time. As YuniKorn sorts priorities solely based on the application and queue, we will assign unique applicationIDs to each Pod and randomly assign different PriorityClasses to the Pods. This approach allows us to observe and analyze the throughput differences between different schedulers.

Result

PriorityClasskube-scheduleryunikornyunikorn plugin mode
makespan993899
throughput(pods/sec)50.5131.650.5
taint-tolerations-result

Summary

The test results reveal that YuniKorn demonstrates the highest throughput across all three tests. Both Kube-scheduler and YuniKorn plugin mode perform comparably.

In tests involving Taint & Tolerations, Affinity & Anti-affinity, and PriorityClass, it's observed that despite numerous parameters and constraints in the deployment process, there's minimal impact on the final throughput and makespan. With the YuniKorn scheduler, a relatively lower throughput was noted, primarily due to the api-server becoming a bottleneck when deploying 5,000 pods simultaneously.