Identify who reports the event
An interrupt lets a device or another event source notify the CPU and transfer control to designated handling. Polling makes the processing side check state to discover whether work has arrived. The comparison concerns notification overhead versus repeated checks when nothing has happened. API names can be misleading: calling a Linux interface named poll or epoll does not necessarily mean that the CPU spins continuously. Those interfaces can put a thread to sleep until readiness or a timeout. Distinguish the application API from a busy-polling implementation and from the device’s notification mechanism. Otherwise an observation about one layer can be incorrectly applied to another layer with different scheduling and waiting behavior.
Calculate the delay of periodic checks
With a simplified check every ten milliseconds, an event arriving just after a check can wait almost ten milliseconds to be noticed. If arrival times are uniformly distributed within that interval, the mean detection delay is about five milliseconds. This calculation excludes service time and scheduling delays. Shortening the interval can improve detection while increasing checks and CPU consumption. A configured interval is not itself a strict end-to-end latency guarantee: the checker still needs execution time, and a busy system can delay it. State the arrival assumption and the measured interval before applying the numerical result. The example explains a trade-off, rather than reporting a benchmark of a particular operating system.
Notification is not immediate completion
Receiving an interrupt does not mean that the application’s request completes at that instant. Masking, priority, deferred work and thread scheduling can all contribute delay. At high event rates, notifying for every small completion can itself create substantial overhead. Interrupt moderation can batch notifications and reduce that cost while potentially adding latency, particularly at low load. Fewer interrupts are therefore not automatically evidence of better service. Observe completions, CPU utilization and response-time distributions together. The relevant question is how the notification policy changes useful work and delay under the workload, rather than whether one counter went down or whether the implementation uses the word interrupt in its description.
Combine the mechanisms under load
Linux NAPI provides a concrete hybrid in networking. In its basic operation, a device notification schedules processing, which handles received packets within a budget. Remaining work can cause further polling; completed work allows coordination back toward notification-driven behavior. This differs both from requiring one interrupt per packet and from unconstrained endless spinning. The budget also helps keep one source of work from monopolizing execution. Exact context and threaded configurations depend on the kernel and driver. Use the documented lifecycle when interpreting counters or implementing a driver, because merely saying that a system uses both mechanisms does not explain when notification is suppressed, how work is rescheduled or how completion is recognized.
Choose from workload and latency requirements
Rare events can favor notification because repeated idle checks consume resources without completing useful work. A dedicated environment with a continuous supply of work and very short latency requirements may instead consider busy polling. Its reserved CPU time, power consumption and impact on other tasks remain part of the cost. Compare low load, ordinary load and bursts, retaining idle intervals rather than hiding them from throughput measurements. Explain detection, processing and return to waiting as a complete cycle. That description makes the choice more meaningful than categorizing the system by one mechanism’s name while overlooking hybrid behavior or the conditions under which a favorable measurement was obtained.