Explore a 15-minute conference talk from USENIX ATC '21 that introduces Faastlane, an innovative approach to accelerating Function-as-a-Service (FaaS) workflows. Dive into the challenges of contemporary FaaS platforms and discover how Faastlane minimizes function interaction latency by executing functions as threads within a single container process. Learn about the implementation of lightweight thread-level isolation domains using Intel Memory Protection Keys for workflows handling sensitive data. Understand how Faastlane dynamically identifies parallelism opportunities and employs process forking or container spawning to execute parallel functions concurrently. Examine the implementation of Faastlane on Apache OpenWhisk and its impressive performance improvements, including up to 15X acceleration of workflow instances and a 99.95% reduction in function interaction latency compared to standard OpenWhisk.
Faastlane - Accelerating Function-as-a-Service Workflows
Overview
Syllabus
Intro
Overview of Function-as-a-Service (FaaS)
Faas workflows: Orchestrating multiple functions
Function interactions in a workflow instance
Faastlane: Running functions in threads
Challenge 1: Protect function's private memory
Healthcare analytics: A case for isolation
Solution 1: Lightweight isolation with Intel MPK
Challenge 2: No thread-level parallelism
FINRA: Parallel functions in a workflow instance
Solution 2: Dynamically switch to processes
Solution 2: Smart scaling with containers
Faastlane: Putting it all together
Function interaction latency on Faas platforms
Request throughput on Faas platforms
Conclusion
Taught by
USENIX
