MemFigLess
An input-aware ensemble learning method — Multi-output Random Forest — for dynamic memory configuration of serverless functions. Achieves 57–87% cost savings and 54–82% memory allocation reduction vs. static baselines on AWS Lambda, evaluated against COSE, Parrotfish, and AWS Power Tuning.
Performance
Evaluated against COSE, Parrotfish, and AWS Power Tuning across diverse serverless function benchmarks.
MAPE Control Loop
Monitor — Offline Training
Profiles serverless functions across diverse input sizes to build a labelled dataset of input characteristics, execution times, and memory consumption. Data stored in AWS DynamoDB and S3.
Analyse — Random Forest Model
Multi-output Random Forest Regressor learns input-to-resource correlations, jointly predicting execution time and memory utilisation for any unseen input payload.
Plan — Constraint Optimisation
Given model predictions, solves a constraint optimisation problem to select the minimum-cost memory configuration that satisfies execution time and SLA constraints.
Execute — Dynamic Configuration
Reconfigures AWS Lambda memory allocation per invocation via the AWS API before function execution. Periodic model retraining via AWS Step Functions maintains accuracy under workload drift.
Benchmark Functions
Mathematical Functions
Matrix multiplication, linear algebra (Linpack), and cryptographic operations (pyaes). Up to 73% resource savings.
Graph Algorithms
Minimum spanning tree, breadth-first search, and PageRank. Up to 87% cost efficiency.
Web & Media Processing
Dynamic HTML generation and template rendering workloads with significant performance improvements.
Publication
Input-Based Ensemble-Learning Method for Dynamic Memory Configuration of Serverless Computing Functions
Agarwal, S. et al. — IEEE CS Press · Sharjah, UAE · DOI: 10.1109/UCC62667.2024
View on IEEE Xplore