Making Serverless Pay-For-Use a Reality with Leopard
summary
Shows that current serverless billing overcharges by charging execution time times a static memory cap, and introduces NPFU billing model that adds CPU caps, preemptable memory, and a cheaper spot‑CPU tier. Builds Leopard on OpenLambda, modifies Linux CFS with a new CPU reserve CPU set cgroup interface, moves OOM decisions to user‑space evictor, and changes admission and cluster schedulers. Evaluation reports >2× provider throughput and lower cost for nearly every invocation.
AI-written summary · about
chapters (6)
- 0:00Serverless billing challenges
- 3:34Nearly pay-for-use model
- 8:41Leopard platform requirements
- 11:09Linux CPU limits
- 14:46Leopard performance
- 17:36Conclusion
related talks
25mCF Serverless: CloudFoundry as a Platform for Serverless Computing
58mServerless Under the Hood
43mDissecting serverless runtimes - from functions to databases
40mWorking with AWS Lambda without Breaking the Bank
17mCloud Computing Made Better: An AWS View on Serverless Computing
42mFaaS or not to FaaS: Visible & invisible benefits of the Serverless paradigm
40mThe State of Serverless Computing
46mServerless: Making Sense with a Cloud Paradox