Ethernet Interrupt Moderation Explained: Latency, CPU Load, and Throughput

A network adapter can notify the host processor when packets arrive or transmit work completes. At high packet rates, handling an interrupt for every event would create substantial CPU and cache overhead. Interrupt moderation reduces that overhead by allowing the adapter to wait for additional packets or a timeout before generating an interrupt.

Microsoft documents interrupt moderation as an NDIS capability and notes that supporting miniport drivers should normally enable it by default. The exact packet count, timeout, or moderation algorithm is hardware- and driver-dependent, so a setting with the same label can behave differently across adapters.

Interrupt handling is not free. Microsoft notes that interrupt latency and cache effects add networking overhead, and that processing several packets per interrupt can reduce this cost under heavy network traffic or heavy system load. Its network-adapter performance guidance reports throughput improvements in some network-intensive workloads when moderation is tuned, while warning that response time can suffer.

Intel describes the same mechanism for its Ethernet adapters: higher interrupt-throttle settings reduce the interrupt rate and can improve CPU performance. That does not mean maximum moderation is universally fastest; it means the system spends less CPU time reacting to individual network events.

The latency side of the tradeoff is straightforward: if the adapter intentionally waits before interrupting the host, an individual packet can spend extra time waiting for service. Microsoft explicitly notes that interrupt moderation can increase the perceived round-trip time for a particular packet.

Intel likewise says a higher interrupt moderation rate means more driver latency when handling packets and documents disabling moderation as useful in some low-latency environments. The cost is potentially much higher CPU utilization because the system must service more interrupts.

A lightly loaded latency-sensitive flow and a saturated high-throughput transfer do not have the same priorities. Intel's current Ethernet guidance exposes settings such as Adaptive, Extreme, High, Medium, Low, Minimal and Off on supported adapters, and says the default is optimal for common network and system configurations.

Windows also defines adaptive interrupt moderation for supported virtual ports, where the miniport driver can adjust the moderation rate according to traffic patterns. Low, medium and high are driver-defined levels rather than universal timing values.

Interrupt moderation controls how frequently network events cause host interrupts. Receive Side Scaling controls how receive processing can be distributed across processors. Checksum and segmentation offloads move particular packet-processing work between CPU and NIC. Link speed describes the physical or negotiated data rate. These features interact with CPU cost and latency, but they are not interchangeable.

A 2.5 GbE or 10 GbE link does not inherently require interrupt moderation to be disabled, and disabling it does not make the link itself faster. Faster links can simply create enough packet-processing work that efficient interrupt handling becomes more important.

For ordinary desktop use, leaving the vendor default or adaptive mode is the sensible baseline. If you have a specific low-latency workload, compare the same traffic pattern with moderation enabled and disabled while watching latency distribution, throughput, CPU utilization and packet loss or drops. For sustained high-throughput workloads, CPU efficiency may matter more than shaving a small amount of interrupt-service delay.

Avoid copying a fixed interrupt-throttle value from another NIC. Intel warns that changing the setting should be validated against the desired network-performance outcome, and Microsoft leaves the actual moderation algorithm to the hardware vendor. Adapter model, driver, queue configuration, traffic pattern and CPU headroom all affect the result.