Capacity planning extends past servers and storage to the network connecting them — it's just another finite resource worth measuring. Networks are plumbing: when healthy, data flows; when not, everything halts. Practical measurement points: - **Devices have ceilings beyond bandwidth.** Routers and switches expose bytes/packets in-out per second via SNMP, but also CPU usage and current sessions. Being well below your bandwidth limit doesn't mean you're not nearing a device's *CPU* ceiling — alert on those too. - **Reconcile against the bill.** In hosted environments network is metered and usage data can be hard to get; aggregate your outward-facing server network metrics and compare them to the provider's invoice as a sanity check. - **Place hosts by network demand.** Owning your racks, divide hosts across switches by the bandwidth each role needs (Flickr spread photo-cache servers thin so each had enough). Use an SNMP-aware grapher (e.g. MRTG) for device history. --- *Source: [[The Art of Capacity Planning]] (John Allspaw, O'Reilly 2008) — Ch 3 — Measurement: Units of Capacity*