WoSC Paper Accepted: Revisiting Pay-per-Use Models in Serverless Applications

The International Workshop on Serverless Computing (WoSC) has become an established forum to discuss new serverless technologies, ranging from application development to runtime concerns. Despite FaaS having raised the bar for not being billed for many hours of idle server processes, we are constantly thinking from an application engineering perspective into how the pay-per-use models could become even more fine-grained and fair. Thus, we are happy to announce that an appropriate autotuning approach has been accepted for being presented at WoSC6 in December 2020.

For several years, we have looked at the overpricing problem from a spatial and a temporal perspective. Spatially, we experience that more memory is allocated than often necessary on two levels – first, dynamic memory needs are not taken into account, as FaaS providers bill by a statically configured limit, and second, due to coarse-grained memory limits, an overallocation is often required even when precise memory limits are known to be lower. Temporally, the first problem does not exist – when a cloud function terminates, it is indeed no longer billed. But the second problem still exists to some degree – even then, the microbilling periods are often leading to what constitutes idle time (even if some providers are now finally going below 100ms periods). There is also an ecological angle to both perspectives in combination – if we reduce the resource consumption over time, less computing needs to be performed overall, and hence the savings could benefit both the FaaS providers and the application providers as well as the environment.

In the presentation, we will focus on the spatial perspective from a purely technical angle. We assume that it is hard to get reliable memory consumption profiles over time for applications due to many influencing factors. But we also assume that eventually such profiles will become available (if we look at the good work conducted in performance engineering for instance). Once we have such profiles, we can request proactive and predictive scaling from the underlying infrastructure.

If you are interested in the details, join us at WoSC6!

Leave a Reply

Your email address will not be published. Required fields are marked *