AeroFlow:Accelerating Serverless Workflows through Serialization-Free WebAssembly Execution

Abstract

Serverless workflow systems orchestrate directed acyclic graphs (DAGs) of short-lived functions to build complex data-processing applications. Conventional designs rely on external storage services to transfer intermediate state between dependent functions, incurring substantial serialization and deserialization overhead on each data-dependency edge. These costs accumulate along the workflow’s critical path and grow super-linearly under concurrency as parallel functions competing for shared storage induce severe resource contention. We present AeroFlow, a low-latency serverless workflow engine that eliminates serialization overhead from inter-function state transfer. AeroFlow uses lightweight, fast-startup WebAssembly sandboxes that intercept and redirect function state accesses at the host-function boundary without requiring changes to user application code. For co-located functions, this replaces storage-mediated communication with serialization-free shared-memory data exchange, yielding notable per-edge transfer latency reduction. A location-aware mixed-integer programming scheduler further optimizes function placement and data routing, co-locating data-intensive function pairs to maximize the fraction of edges served via the fast local path. On both scientific workflow benchmarks and real-world serverless workloads, AeroFlow achieves an end-to-end performance improvement of 2.4–12.4 × over state-of-the-art systems at moderate scales. At a scale of 250 functions, the speedup reaches 7.8 × , as baseline systems exhibit super-linear performance degradation; at a scale of 500 functions, AeroFlow is the only system that successfully completes execution.