Abstract: Serverless computing platforms, such as Function-as-a-Service (FaaS), have been gaining popularity and are considered by some as the future of cloud computing. Serverless applications benefit from highly elastic resources that are fast to instantiate, scale to thousands of instances and are based on fine-grained resource accounting. These attractive resources are bundled with the FaaS programming model, based on restricted event-triggered functions that applications can compose into complex workflows. This can provide benefits such as autoscaling and reduced management overhead. However, limitations such as the restricted function environment, limited execution duration, or restricted networking can be obstacles for some applications, such as distributed data analytics. Furthermore, these resources are not directly available to classic datacenter applications based on the network-of-hosts programming model.
This thesis proposes to unbundle the serverless resources from the programming model, and demonstrates that it is possible to use existing publicly available FaaS infrastructure to transparently provide serverless resources to datacenter applications. Boxer system is developed that, transparently to the application, provides the classic network-of-hosts model based on publicly available FaaS resources to enable serverless datacenter applications. It shows how such serverless datacenter applications can benefit from augmented ephemeral elasticity and can be instantiated on per-request granularity.
It is demonstrated that unmodified long-running datacenter applications can be transparently augmented with ephemeral elasticity based on FaaS, reducing the overprovisioning needed as application resources can scale faster in response to demands, such as bursty workloads. An unmodified microservice benchmark shows that its applications, when using Boxer-provided ephemeral elasticity, scale as quickly as with overprovisioned virtual machines. It is also shown that ephemeral elasticity can significantly reduce the failure recovery times in unmodified distributed datacenter applications as new resources can be quickly provisioned compared to virtual machines. Furthermore, serverless datacenter applications can also be instantiated on per-request granularity, in particular, it is shown how distributed off-the-shelf data analytics systems can be instantiated on per-query granularity using publicly available FaaS resources instead of submitting queries to a long-running pre-selected and configured data processing system.
@PhdThesis{wawrzoniakPhd2024,
title = {Serverless Datacenter Applications},
author = {Michael Wawrzoniak},
school = {ETH Zurich},
year = {2024}
}