Packet Coalescing & Interrupt Moderation — Architecture & Performance

Overview

Packet coalescing and interrupt moderation are techniques used in network interfaces and operating systems to reduce CPU overhead, improve throughput, and optimize power consumption. They work by aggregating multiple packets or events before delivering them to the host CPU, thereby reducing the number of interrupts and context switches.

These mechanisms are essential in high‑speed networks, virtualized environments, and mobile systems, where interrupt storms and per‑packet processing overhead can severely limit performance.

Goals of Coalescing and Moderation

These mechanisms aim to:

• reduce CPU interrupt rate
• improve throughput under high load
• lower power consumption
• amortize per‑packet processing overhead
• maintain acceptable latency for interactive traffic
• balance efficiency and responsiveness

The challenge is to optimize performance without harming latency‑sensitive applications.

Architectural Principles

Interrupt Cost

Each interrupt triggers:

• a context switch
• cache pollution
• driver processing
• potential wake‑ups from low‑power states

Reducing interrupt frequency improves efficiency.

Batch Processing

Processing packets in batches amortizes overhead across multiple packets.

Batching improves throughput but increases latency.

Adaptive Behavior

Modern NICs dynamically adjust coalescing parameters based on:

• traffic rate
• packet size
• CPU load
• latency requirements

Adaptive moderation provides the best balance across workloads.

Packet Coalescing

Concept

Packet coalescing aggregates multiple received packets into a single buffer or descriptor before delivering them to the OS.

Benefits:

• fewer DMA operations
• reduced descriptor processing
• improved cache locality

Coalescing is especially effective for small packets.

Receive‑Side Coalescing (RSC)

RSC combines multiple TCP segments into a larger, reassembled segment before passing it to the OS.

Benefits:

• reduces per‑packet overhead
• improves throughput for small TCP segments
• reduces CPU load

Used in modern NICs and OS network stacks.

Large Receive Offload (LRO)

LRO aggregates incoming packets belonging to the same flow.

Differences from RSC:

• LRO is more aggressive
• may break TCP semantics (e.g., out‑of‑order handling)
• not suitable for routing or forwarding devices

Used primarily in servers and end hosts.

Transmit‑Side Coalescing

On the transmit path, NICs may combine small packets into larger frames or batch descriptors.

Benefits:

• reduces DMA overhead
• improves throughput
• lowers CPU usage

Often integrated with TSO (TCP Segmentation Offload).

Interrupt Moderation

Concept

Interrupt moderation delays or batches interrupts so that multiple packets trigger a single interrupt.

Moderation parameters include:

• interrupt delay (timer‑based)
• interrupt threshold (packet count)
• adaptive algorithms

Moderation reduces interrupt storms under high load.

Timer‑Based Moderation

Interrupts are generated only after a timer expires.

Pros: predictable
Cons: increases latency for small bursts

Threshold‑Based Moderation

Interrupts are generated after a certain number of packets or bytes.

Pros: efficient under high load
Cons: may delay small flows

Adaptive Interrupt Moderation (AIM)

NICs dynamically adjust moderation based on traffic patterns.

Benefits:

• low latency under light load
• high throughput under heavy load
• reduced CPU usage

AIM is widely used in modern NICs.

Interaction with Offload Engines

TSO (TCP Segmentation Offload)

TSO reduces transmit overhead by offloading segmentation to the NIC.

Coalescing complements TSO by reducing receive overhead.

GRO (Generic Receive Offload)

GRO is a software‑based coalescing mechanism in the OS.

Benefits:

• flexible
• protocol‑aware
• works with virtualized environments

GRO is widely used in Linux.

GSO (Generic Segmentation Offload)

GSO performs segmentation in software before NIC transmission.

Useful when hardware TSO is unavailable.

Performance Considerations

Latency

Coalescing and moderation increase latency because packets wait before being processed.

Latency‑sensitive applications (VoIP, gaming, HFT) may require:

• minimal coalescing
• low interrupt delay
• priority queues

Throughput

Batching improves throughput by:

• reducing per‑packet overhead
• improving cache locality
• reducing interrupt frequency

High‑speed NICs rely heavily on coalescing to sustain line rate.

CPU Utilization

Moderation reduces CPU load by:

• lowering interrupt rate
• reducing context switches
• improving batching efficiency

Critical for servers and virtualized environments.

Power Consumption

Mobile and embedded systems use coalescing to:

• reduce wake‑ups
• extend battery life
• maintain low‑power states longer

Use Cases in Modern Systems

Data Center Servers

High‑speed NICs (25G/50G/100G/200G) rely on:

• RSC/GRO
• TSO/GSO
• adaptive interrupt moderation

These mechanisms are essential for sustaining high throughput.

Virtualized Environments

Hypervisors and virtual NICs use:

• software coalescing
• batching
• interrupt moderation

These reduce VM exits and improve scalability.

Mobile and IoT Devices

Coalescing reduces:

• power consumption
• CPU wake‑ups
• radio transmission overhead

Used in Wi‑Fi, LTE/5G modems, and low‑power NICs.

High‑Speed Interconnects

PCIe, CXL, and NoCs use:

• event coalescing
• credit batching
• interrupt moderation

These ensure deterministic performance and reduce overhead.

Comparison of Coalescing and Moderation Techniques

Technique	Latency	Throughput	CPU Load	Typical Use
Timer‑Based Moderation	Medium	High	Low	NICs, OS
Threshold‑Based Moderation	Medium	High	Low	High‑speed NICs
Adaptive Moderation	Low-Medium	Very High	Very Low	Modern NICs
RSC/GRO	Medium	Very High	Low	Servers
LRO	Medium	High	Low	End hosts
TSO/GSO	Low	Very High	Very Low	Servers, virtualization

Design Tradeoffs

• Latency vs throughput — batching improves throughput but increases delay.
• CPU load vs responsiveness — fewer interrupts reduce CPU usage but delay processing.
• Hardware vs software — hardware is faster; software is more flexible.
• Power vs performance — coalescing saves power but may harm latency.
• Flow granularity vs batching — fine‑grained control reduces batching efficiency.

Related Pages

• Generic Packet Parsing — Architecture & Pipelines
• Packet Scheduling — Architecture & Algorithms
• Queue Management & Congestion Control — Architecture & Algorithms
• Traffic Prioritization & Service Classes — Architecture & Use Cases
• Link Layer Reliability Mechanisms — Architecture & Design Patterns

Summary

Packet coalescing and interrupt moderation reduce CPU overhead, improve throughput, and optimize power consumption by batching packets and events before delivering them to the host. Through RSC, GRO, TSO, adaptive moderation, and batching techniques, modern systems achieve high performance across data centers, virtualized environments, mobile devices, and high‑speed interconnects.