GRANNY: Granular Management of Compute-Intensive Applications in the Cloud
summary
A distributed runtime called GRANNY uses a WebAssembly-based execution abstraction to run one-vCPU Granules that offer either thread (OpenMP) or process (MPI) semantics, enabling vertical scaling by adding Granules and horizontal consolidation via snapshot-and-migrate. Snapshotting relies on WebAssembly’s linear memory and control-point APIs, and experiments with 100 LAMMPS instances on 32 VMs (8 vCPUs each) leave 5% headroom; results show reduces makespan by up to 60% for OpenMP and reduces fragmentation by up to 25% for MPI.
AI-written summary · about
chapters (9)
- 0:00GRANNY project intro
- 1:20Compute-intensive cloud apps
- 3:06Cloud resource allocation challenges
- 5:14GRANNY runtime for dynamic management
- 8:28Granules execution model with WebAssembly
- 10:56Safe granule management
- 12:43Granule migration across VMs
- 13:47GRANNY policies and experiments
- 16:40Conclusion and future work
related talks
11mLightning Talk: MPIWasm: Executing WebAssembly on HPC Systems
16mDisaggregation and the Application
16mPreview: Cloud Computing
49mGraalVM + Wasm: Fast, Efficient, Portable Apps
17mHierarchical Integration of WebAssembly in Serverless for Efficiency and Interoperability
33mEfficient and Cross-Platform LLM Inference in the Heterogenous Cloud
30mA Greener, Cost-Effective Cloud with Serverless WebAssembly
16mWasmgrind: Towards Dynamic Concurrency Analysis for Multithreaded WebAssembly