HarvestContainers safeguards latency‑sensitive pods while reusing idle cores – The system isolates latency‑critical containers from interference and concurrently runs throughput‑oriented containers on their spare CPU capacity. It protects against all sources of interference, including networking interrupts. No modifications to applications or operating system are required. Evaluation integrates it with Kubernetes. [0]
Dynamic core‑harvesting determines safe CPU allocation in real time – HarvestContainers continuously assesses each container’s CPU usage to decide how many unused cores can be allocated to secondary workloads. The approach avoids offline profiling and static provisioning. It ensures that latency targets remain met while maximizing resource utilization. [0]
Experimental results show up to 75% of unused cores can be shared – In tests, latency‑sensitive containers with microsecond‑scale service‑level objectives retained tail latency within 4% of their standalone performance while sharing three‑quarters of their idle cores. This demonstrates substantial efficiency gains without compromising quality of service. [0]
Existing resource controls cannot share cores without latency penalties – Current mechanisms either over‑provision CPU or require application changes, and they ignore interference from container networking interrupts. Prior research on performance isolation demands code rewrites and offline analysis, making it unsuitable for modern container environments. [0]
Kubernetes over‑provisions CPU to meet 99th‑percentile latency under peak load – Cloud services experience bursty traffic, causing CPU usage variability that leaves cores idle between spikes. This over‑provisioning creates the opportunity HarvestContainers exploits. [0]