Optimizing Latency Sensitive Applications for Amazon's Public Cloud Platform

Operating Latency Sensitive Applications on Public Serverless Edge Cloud Platforms

“…The system is capable of composing, deploying and dynamically reoptimizing IoT applications operated on serverless resources. We note, that the first and basic version of the system, without any support for the edge, was introduced in [7] and [29]. In the former, we focused on the optimization layer, while in the latter, we investigated deployment tasks.…”

“…In our previous works [7], [28], we provided a comprehensive performance study of delay characteristics of AWS FaaS and CaaS offerings, based on short-and long-term experiments. Here, we give a summary on them focusing on our main findings with regards to AWS Lambda.…”

“…For the runtime environment, we only consider single-threaded serverless functions. However, our models can be extended to use containers [7] or to support multi-threaded functions by using explicit function execution profiles. Runtime flavors Φ are specified by their offered vCPU fraction n c .…”

“…Since we aim to split up trees explicitly into chains and the partition groups are bounded by the latency limits of service-wide critical paths in contrast to the locally verifiable group size or count limits, the aforementioned methods cannot be applied directly to our problem. By extending our prior algorithm designed for public clouds [7], we propose a heuristic approach for cost-efficient and latency-constrained partitioning of trees into chains on cloud and edge resources, called Chain-based Tree-Partitioning (CTP).…”

“…More precisely, we focus on serverless architectures and the Function as a Service (FaaS) cloud computing model where the microservice based application is built from isolated functions which are deployed and scaled separately by the cloud platform. In our previous work [7], we proposed a novel mechanism to optimize the software "layout", i.e., to minimize the deployment costs, in central cloud environment, e.g., in a given AWS region, while meeting the average latency constraints defined on the application. A dedicated component is responsible for composing the service by selecting the preferred building blocks, such as runtime flavors (defining the amount of resources to be assigned) and data stores, and the optimal grouping of constituent functions and libraries which are packaged into respective FaaS platform artifacts.…”