Performance Monitoring, Troubleshooting, and Optimization
Troubleshooting questions give a symptom - high latency, low IOPS, a stalled migration - and ask for the first diagnostic step or the most likely cause. Know which metric lives where and which tool to reach for.
Where to look
- Prism Analysis: cluster IOPS, throughput, latency (hypervisor versus storage/controller), CPU ready time, memory pressure and per-VM charts.
- nCLI and aCLI from the CVM for object-level detail; NCC for health checks and known-issue detection; CVM logs under /home/nutanix/data/logs.
- Prism alerts and Pulse/Insights for pattern-based findings across the estate.
Common root causes
- Network: wrong bond mode, missing LACP, MTU mismatch, one dead uplink.
- CVM starvation: too few vCPUs or too little RAM on the CVM, throttling all I/O for that host.
- Contention: snapshot deletion, Curator scans, replication or backups overlapping production peaks.
- Lost data locality after a host or CVM restart, producing remote reads.
Worked example
Users report latency every night at 02:00. Analysis shows storage controller latency spiking while IOPS stays normal; the same window contains a backup job and a protection-policy snapshot burst. Reasoning: throughput is not the constraint, so queueing is inside the DSF. Stagger the backup start and snapshot schedules, confirm OpLog drain returns to baseline, and re-check with NCC that no disk is degraded.
Optimization
Enable compression cluster-wide, enable dedup where data is cold and redundant, use EC-X for cold data, right-size vCPU (CPU ready time is the tell), keep CVM resources unreserved, and align affinity rules with the physical layout.
Traps
- Reading only VM-level latency and missing the storage controller.
- Changing several variables at once so the fix cannot be attributed.
- Adding vCPUs to a VM that is already waiting on CPU ready.