Performant &
Highly Available & Cost Efficient
Spot Instances for Kubernetes
KubePACS is a Kubernetes-native spot instance provisioning system that constructs node pools optimized for both cost and performance while guaranteeing high availability. It redefines cloud resource allocation by maximizing performance-per-dollar utilizing real-time cloud datasets.
Key Challenges
Performance Heterogeneity
Cloud instances often vary vastly in computing power, yet traditional tools scale nodes based merely on static CPU/memory requirements without accounting for benchmark performance variation.
Inadequate Single-Node Metrics
Existing approaches infer cluster-level availability from single-node placement scores, exposing multi-node workloads to high interruption risks. KubePACS utilizes Multi-Node SPS.
Lack of Workload Awareness
Spot price alone decoupled from hardware capability makes it challenging to select ideal nodes for specialized workloads (e.g., Disk/Network I/O bound tasks).
How KubePACS Works
Multi-Objective Optimization Pipeline
KubePACS solves the complex multi-objective optimization problem to build the perfect cluster. It seamlessly integrates real-time cloud datasets into a three-stage automated pipeline.
Metric Preprocessor
Aggregates spot prices, benchmark scores (CoreMark), and multi-node SPS, scaling performance for workload-specific I/O capabilities.
ILP Node Selection Solver
Formulates an Integer Linear Programming (ILP) problem to balance cost performance and avoid over-allocation while satisfying demand.
GSS Optimizer
Iteratively applies the Golden Section Search (GSS) algorithm to efficiently identify the optimal cost-performance trade-off hyperparameter (α).