Faastlane is a FaaS tool that accelerates workflows by minimizing function interaction latency. The key contributor to this latency is the transient state that needs to be shared amongst functions within a single workflow instance. Arguments passed between functions in a workflow instance are examples of such transient state. Faastlane strives to execute functions of a workflow instance as separate threads within a single process. Functions communicate simply via load/store instructions, thus minimizing function interaction latency. The shared virtual address space of the encompassing process provides the simplest form of namespace for sharing state — it does not necessitate a new API. To ensure security, each thread executes in its own memory domain and each domain is protected using Intel's MPK (Memory Protection Keys).
Faastlane appeared at the USENIX Annual Technical Conference, 2021. For more details, please refer https://www.usenix.org/conference/atc21/presentation/kotni
This folder contains a custom memory allocator that allocates from different memory pools based on the thread id of the incoming memory request. Each thread is assigned a protection key (pkey) and all pages serving requests for that thread are protected with the corresponding pkey.
This folder contains Faastlane's client-side composer. The composer generates an orchestration function that provides an illusion of single function to a FaaS platform.
Usage: python3.6 generator.py workflow.json
Virtualenv that contains a working version of the memory allocator module.
Util scripts to test memory allocator