Χρήση διαφορικών στιγμιοτύπων για την βελτιστοποίηση εκκίνησης Serverless workloads σε ARM επεξεργαστές

Abstract

Serverless architecture is a modern cloud-computing paradigm. In collaboration with Function-as-a-Service applications, it promises high utilization of the resources provided by data centers through the allocation of resources on demand. An interesting application of the Serverless architecture is the use of Internet of Things devices, which are constantly increasing. A major limitation of the Serverless architecture is the delay resulting from the activation of the underlying infrastructure (microVM, container, unikernel) of each Function-as-a-Service application, in order to serve a request. The cold-start process, as it is called, refers to cases where our application in the Serverless environment has not been invoked recently and thus, the virtual machine in which it is running is not in an idle state, nor activated. So when an invocation to that function is made, the microVM must be set up entirely from scratch, which adds a delay to the service of the request. A potential solution that has been studied thoroughly by the scientific community is the use of snapshots, through which we can store the current state of the underlying infrastructure and restore it (when needed) with negligible overhead. Previous research work has shown that we can achieve a reduction in cold-start latency through different snapshotting techniques. In this paper we studied the SnapFaas protocol which uses new techniques to implement snapshots and which promises fast End-to-End request servings. Based on the growing trend of IoT devices that primarily use ARM processors, we inspected the porting of SnapFaas system to ARM-based computing systems. We will analyze the changes needed to achieve this and also quote some code snippets from the implementation. Additionally, we will study how snapshots are taken by SnapFaas, capturing as efficiently as possible the memory and the various devices attached to the microVM. Accordingly, we will examine how these snapshots are used to restore the microVM, both in the memory part, as well as the part of the connected devices.