Design And Implementation Of A Serverless Cloud Computing Platform Based On Knativ

As a new paradigm of cloud services,serverless computing has attracted extensive attention from industry and academia since its inception.Serverless has two major advantages.One is fine-grained on-demand loading and pay-as-you-go,and the other is complete O&M-free.This requires the serverless cloud computing platform to have a high degree of resource autonomy,such as automatic scaling of resources.A good auto-scaling strategy can improve platform throughput and reduce service response time.But at the same time,due to the technical characteristics of serverless function instances shrinking to 0,the function cold start problem caused by it is also an inherent performance problem that the platform needs to consider.Therefore,in order to realize a serverless cloud computing platform and improve the overall service quality level of the platform,this paper mainly completes the following work:(1)Based on the open source framework Knative,a serverless cloud computing platform that manages the entire life cycle of cloud functions is designed and implemented,including the design and development of the platform’s overall technical architecture and functional module structure.At the same time,according to the serverless task load characteristics and related research,this paper constructs a set of load benchmark task set Function Bench including 5 task categories and 26 cloud functions.(2)A serverless platform automatic scaling strategy PS-Scale based on the deep reinforcement learning algorithm PPO is proposed.This strategy models the auto-scaling problem in the Serverless scenario as a Markov decision process,treats cluster resource status,load resource consumption,and dynamic configuration of functions as a continuous space for control,and introduces the PPO algorithm to learn to find solutions in the current solution space The best auto-scaling configuration builds a load-adaptive serverless platform autoscaling strategy.Finally,a concurrency experiment in a real cluster environment is carried out on the experimental platform based on Knative.The results show that compared with the existing automatic scaling strategy,PS-Scale has a certain improvement in the average throughput and response delay indicators of the platform.(3)A serverless cloud function cold start optimization strategy FS-Warm based on function set is proposed.First of all,based on the Azure Function trace public data set and the existing advanced scheduling strategy adaptive hybrid histogram strategy(HHP),this paper conducts data analysis and finds the key defects of the existing strategy.Then use the analysis conclusion to guide the improvement of the strategy.Based on the FP-Growth algorithm,a function set whose granularity is between the function and the application is generated to participate in the scheduling,and the memory is saved as much as possible under the premise of making full use of the association relationship between cloud functions to reduce cold starts resource.Finally,simulation experiments on real data sets show that FS-Warm can achieve better results in reducing the function cold start rate and balancing memory resource consumption.