Capacity, Performance, and Health Management
This domain is about reading the array rather than configuring it: what consumes capacity, where latency originates, and how health is surfaced.
Capacity accounting
- Logical capacity is what hosts were given; physical capacity is what was actually written. Thin provisioning, zero-block detection, and inline deduplication and compression create the gap reported as the data reduction ratio.
- Snapshot space comes from the same pool as the volume family. A generous retention rule on a high-change-rate volume can consume more than the volume itself.
- Data reduction can be controlled per volume in current releases; check the setting before assuming a ratio applies to already-compressed or encrypted workloads.
- Capacity alerts and CloudIQ forecasting buy lead time. Expansion typically means adding an enclosure to the existing appliance rather than mixing drive types.
Performance reading
- Front-end metrics (host-facing IOPS, bandwidth, latency) and back-end drive metrics tell different stories. High front-end latency with low back-end utilisation points at the network, the host, or pathing; both high points at the array.
- Check node balance. Workload concentrated on volumes preferring one node leaves node A busy and node B idle; volume mobility moves a volume to the other node.
- IO size explains IOPS-versus-bandwidth inversions: 4 KB random and 256 KB sequential figures are not comparable.
Health
- The health score and alert list aggregate component state; high drive wear, a failed fan, or a degraded NVRAM battery surfaces there first.
- Run the system health check before any maintenance window and confirm SupportAssist so failures raise cases automatically.
Worked example
Users report slowness; the array shows 40 percent front-end utilisation but 8 ms latency, with node A at 95 percent and node B at 5 percent. Reasoning: the constraint is imbalance, not raw capability. Move the busiest volumes to node B and re-measure. Only after balance is corrected should you consider adding an appliance, because a new appliance does not fix a skewed preferred-node layout.
Traps
- Comparing IOPS figures captured at different IO sizes.
- Treating logical used as physical consumption, or expecting deduplication on pre-compressed data.
- Blaming disks for latency before checking front-end, switch, and host statistics.
- Ignoring snapshot growth caused by an over-generous retention policy.