On shared VPS, other tenants can affect your CPU and I/O. Monitor steal time and disk latency. If contention is persistent, move to a larger VPS, dedicated CPU, or bare metal. Tier-III providers often cap oversubscription.
What is contention
- Noisy neighbors: On shared hardware, one tenant's heavy CPU, memory, or disk usage can slow others. The hypervisor shares physical cores and I/O; there are no hard guarantees unless you pay for dedicated resources.
- CPU steal time: Time the VM wanted to run but the hypervisor gave the CPU to another VM. High steal (e.g.
vmstat,top"st") indicates CPU contention. Sustained high steal hurts performance. - Disk I/O: Shared storage can be slow when others do heavy I/O. Monitor disk latency (e.g.
iostat, cloud metrics). Spiky or high latency suggests storage contention.
Detection
- Metrics: Track CPU steal, disk read/write latency, and load average. Compare with your app's latency and throughput. Correlate dips with time of day or other tenants' activity if visible.
- Provider tools: Some providers expose host-level or aggregate metrics. Use them to confirm that contention is on the host side rather than your app.
- Escalate: If you have an SLA, open a ticket with evidence (steal time, latency graphs). Ask for a different host or a commitment to reduce oversubscription.
Mitigation
- Short term: Optimize your app (reduce CPU or I/O), add caching, or scale out (more small instances may spread load across hosts). Sometimes a reboot or resize moves you to a less loaded host.
- Long term: Move to a larger VPS with guaranteed resources, a "dedicated CPU" or "burstable" plan with a higher baseline, or bare metal. Tier-III or premium providers often limit oversubscription (e.g. 4:1 or 2:1) and are more predictable.
- Design: For critical workloads, design for contention: stateless app, horizontal scaling, and tolerance for occasional latency spikes. Use queues and timeouts so one slow node does not block the system.
Summary
On shared VPS, other tenants can cause CPU and I/O contention. Monitor steal time and disk latency. If persistent, optimize, escalate, or move to dedicated CPU / bare metal. Choose providers that cap oversubscription for predictable performance.
