Current serverless solutions are primarily designed to run on cloud centric environments. While bringing these solutions to the edge is its further evolution, it introduces new challenges, due to resource constraints, different CPU architectures, network topology and latency. Specifically, when clustering devices at the edge, inter-node latency plays an important role. In this paper, we experimentally examine the impact that latency has on scalablity by measuring the throughput of distributed serverless applications. We deploy Knative over a Kubernetes cluster of nodes and emulate latencies between them to compare the performance of serverless functions when deployed over centralized versus distributed computing sites. The results show how scaling over edge achieves half the throughput as compared to a centralized deployment in the cloud when the processing times are low, but more than two thirds the improved performance of cloud with increased processing delays.

翻译：当前的无服务器解决方案主要设计用于以云为中心的环境。将这些解决方案扩展到边缘是其进一步发展，但由于资源限制、不同的CPU架构、网络拓扑及延迟，带来了新的挑战。具体而言，当在边缘对设备进行集群化时，节点间的延迟起着重要作用。本文通过测量分布式无服务器应用的吞吐量，实验性地考察了延迟对可扩展性的影响。我们在一个Kubernetes节点集群上部署Knative，并模拟节点间的延迟，以比较无服务器函数在集中式计算站点与分布式计算站点上部署时的性能。结果表明，当处理时间较短时，边缘扩展的吞吐量仅为云端集中部署的一半；而当处理延迟增加时，边缘扩展的性能可达到云端改进性能的三分之二以上。