Modern monitoring tools can track thousands of metrics. The challenge isn't getting data; it's knowing which data signals a real problem versus which is just noise. Effective monitoring focuses on the "USE" method: Utilization, Saturation, and Errors.
1. CPU: Load vs. Usage
CPU Usage (%) tells you how busy the processor is right now. But CPU Load Average is often more important. It tells you how many processes are waiting for their turn. A usage of 100% with a low load average means the CPU is busy but keeping up. High usage with High Load means the system is choking and tasks are queuing up.
2. Memory: Available vs. Free
"Free ram is wasted ram." Linux caches frequently accessed files in unused RAM to speed up performance. This means "Free" memory often looks near zero, causing false alarm. The metric to watch is "Available"—this is memory the kernel can instantly free up if applications need it. Ignore "Free"; monitor "Available."
3. Disk I/O: Wait Time
Disk usage (%) is irrelevant for performance; Disk I/O Wait (iowait) is critical. High iowait means the CPU is sitting idle, doing nothing, just waiting for the hard drive to read/write data. This is the #1 silent killer of server performance. If iowait exceeds 10-15%, you likely need faster storage (NVMe) or more RAM to cache data.
4. Application Metrics
Infrastructure metrics are proxies. The ultimate truth lies in application metrics.
- Latency: How long does a request take?
- Error Rate: What percentage of requests return 5xx errors?
- Throughput: Requests per second.
A server with low CPU usage but high application error rates is still broken. Monitor from the user's perspective first, the server's perspective second.