Title Page
ABSTRACT
Contents
Ⅰ. Introduction 10
Ⅱ. Background and Related Work 15
1. Autoscaling mechanism of Kubernetes 15
2. Knative runtime and its auto-scaler 18
Ⅲ. Benchmarking Overview 21
1. Aim and Scope 21
2. Custom metrics related to autoscaling 23
3. Benchmark suites and workload specification 25
Ⅳ. Evaluation and Implications 28
1. Correlation between auto-scaling and the initial resource allocation of pods 30
2. Auto-scaling efficiency depending on scaling threshold 40
3. Auto-scaling behavior of HPA depending on base resource metrics 47
Ⅴ. Selecting a default autoscaling policy: HPA vs KPA 53
1. Comparison of auto scaling performances in HPA and KPA with their default metrics 53
2. Miscellaneous : FaaS deployment breakdown 59
Ⅵ. Conclusion 62
References 63
논문 개요 70
TABLE 1. Detailed specification of deployed functions 27
TABLE 2. Knative's key process to ready a pod 60
FIGURE 1. Typical FaaS architecture and system components utilizing Knative on Kubernetes 18
FIGURE 2. An overview of FaaS environmental setup for function deployment and invocation 25
FIGURE 3. (a) Comparison of the scaling process with different threshold settings (b) Comparison of the Number of [200] responses (c)... 31
FIGURE 4. (a) Comparison of the scaling process with different initial pod resource settings (b)Comparison of the Number of [200] responses... 35
FIGURE 5. (a) Comparison of the scaling process with different threshold settings (b) Comparison of the Number of [200] responses (c)... 43
FIGURE 6. (a) Comparison of the scaling process (b) Comparison of the Number of [200] responses (c) comparison of the overall latency distribution 49
FIGURE 7. (a) Concurrency level of 80 (b) Concurrency level of 600 55
FIGURE 8. (a) Trends in response according to concurrency level(1~100) (b) Trends in SLA of HPA (c) Trends in SLA of KPA 58
FIGURE 9. (a) Trends in response according to concurrency level(100~600) (b) Trends in SLA of HPA (c) Trends in SLA of KPA 58
FIGURE 10. Time spent on pod ready process 61