AI-Powered IEEE ICDCS 2026 · Under Review

Saarthi

An end-to-end intelligent serverless computing platform integrating an ML-based request scheduler, custom gateway, load balancer, and autoscaler. Achieves 1.84× cost reduction and 97.2% SLA compliance benchmarked against AWS Power Tuning and standard Kubernetes HPA.

OpenFaaS

Python

Kubernetes

Prometheus AWS

AWS

Live Demo arXiv Preprint

Results

Performance

1.84×

Cost reduction vs. baseline

97.2%

SLA compliance rate

~0.1 ms

Scheduler inference latency

3

Industry baselines beaten

Benchmarked against AWS Power Tuning, standard HPA, and baseline OpenFaaS.

Design

Architecture

Prediction Layer

Advanced ML models forecast workload patterns and resource requirements ahead of arrival, enabling proactive warm-up and eliminating cold starts under predictable traffic.

Optimisation Layer

Integer Linear Programming (ILP) solver computes mathematically optimal resource allocation, maximising utility subject to SLA and capacity constraints.

Execution Layer

Custom OpenFaaS gateway with idle-first pod selection, real-time request routing, and Kubernetes Custom Resource Definitions (CRDs) for function lifecycle management.

Monitoring Layer

Prometheus-backed multi-dimensional telemetry with cost analytics, per-function SLA tracking, and real-time rebalancing triggers.

Stack

Technical Implementation

Core Technologies

Kubernetes-native deployment with Custom Resource Definitions
OpenFaaS with custom gateway and provider extensions
Integer Linear Programming (ILP) for resource allocation
Prometheus & Grafana for telemetry collection and dashboards
Multi-node Kubernetes clusters (5–20 nodes) on Melbourne Research Cloud (NeCTAR) and AWS

Advanced Capabilities

Idle-first pod selection for minimal cold start overhead
Predictive autoscaling independent of standard HPA
Sub-millisecond scheduling inference at production load
Validated against ICDCS-ready experimental benchmarks

Research

Publication

IEEE ICDCS 2026 · Under Review

Saarthi: An End-to-End Intelligent Platform for Optimising Distributed Serverless Workloads

Agarwal, S. et al. — arXiv:2511.06599

View on arXiv