Novel architecture for mitigating cold start problem in serverless computing

Abstract

Features like no server maintenance, low cost, and elasticity are attracting developers to build applications utilizing the functions-as-a-service (FaaS) architecture on serverless platforms.Researchers have proposed various fixes to overcome the cold start problem, a well-known issue of serverless architecture. In thisthesis,weaddressthecoldstartproblemoftheserverlessplatform. Ourstudy shows that we can not get rid of cold start time because of the on-demand nature of the FaaS model.In addition, developers are adopting FaaS for critical stateful applications to retain the architecture’s benefits. Consequently, the need for an adaptive and intelligent fault-tolerant technique for FaaS architecture has transpired.It has become evident that the classic Retry technique is neither reliable nor efficient for handling such critical applications. Admittedly,mitigatingcoldstartlatencyandachievingfaulttoleranceforserverless has remained an open problem.Therefore, we propose Warm Lambda Equipped Container (WLEC), a container management architecture that utilizes a modified TwoSegmentedLeastRecentlyUsed(S2LRU)structurecalledS2LRU++inspired bytheS2LRU*architectureofPannier.Then,weevaluateWLECinbothAWSand Local VM environments with OpenLambda with six different metrics. Assessing the practical implication and performance, we realize the shortcomings of WLEC and decide to alleviate them.Subsequently, we have developed Advance Warm LambdaEquippedContainer(AdWaLEC)thataddressescoldstartlatencyandfault tolerance with replication-based strategies. Moreover, AdWaLEC redesigns WLEC-one of the cold start mitigation solutions, by improving container replication of template container list utilizing Spearman’s rank correlation coefficient to reduce cold start delay. Next, it also incorporates a reliable fault tolerance scheme for the ubiquitous serverless model called Warm LambdaContainerReplication(WaLCoR)thatleveragescontainer-associatedlogs. Further,wehaveimplementedAdWaLECintheOpenLambdaplatform,evaluated itwithAzureFunctiontracewithsevendifferentmetrics,andcompareditwithfour different approaches. For instance, AdWaLEC shows a 77.23% reduction in cold start frequency, a 27.36% decrease in cold start latency, and a 29.95% increase in invocation per container compared to the baseline of OpenLambda.