Team project · Jan — Apr 2025 · IIT Jodhpur
Cost-efficient Kubernetes orchestration
A custom Kubernetes scheduler and hybrid task autoscaler (HTAS) for GKE that packs microservices and batch jobs onto heterogeneous, preemptible VMs.
- Role
- Cloud architect, Systems engineer
- Timeline
- Jan — Apr 2025
- Type
- Team project, public repo
- Stack
- Kubernetes, GCP, Python, Prometheus
↓28%Compute costComparative GCP billing across identical workloads
+35%UtilisationAverage CPU and memory efficiency on the test cluster
10+Services orchestratedMicroservices and batch workloads balanced across nodes
01Context
High GKE costs from inefficient default scheduling created the need for custom orchestration tuned for mixed workloads and preemptible resources.
02Approach
A modular Kubernetes extension built on CRDs, with autoscaling driven by real-time resource profiling and integrated with GCP's managed cluster APIs.
- Built the scheduler and autoscaler (HTAS) on CRDs with event-driven scaling for GKE clusters.
- Designed a Resource Profiler that characterises workloads, and a Task Packer that places them with BFD/TBFD bin-packing.
- Benchmarked synthetic and real workloads, tuning placement heuristics for mixed VM sizes.
03Impact
A cost-conscious orchestration platform with measurable infrastructure savings and better utilisation, informing further cloud infrastructure research at IIT Jodhpur.
- 28% lower cost and 35% higher utilisation than the default GKE scheduler.
- Fault-tolerant scheduling that stays correct through preemptible VM churn.
- Per-pod telemetry in Prometheus for real-time scaling decisions.