Why is Kafka so Fast? Part 2

Key Concepts:

• Zero Copy: A technique to eliminate unnecessary data copying between disk, OS cache, application memory, and network interface card (NIC) buffer, improving data transfer efficiency.
• OS Cache: The operating system's buffer for frequently accessed data from disk, enabling faster retrieval.
• Socket Buffer: A memory buffer associated with a network socket, used to hold data being sent or received over the network.
• Network Interface Card (NIC) Buffer: A memory buffer on the network card that holds data before it's transmitted over the network.
• DMA (Direct Memory Access): A hardware feature that allows devices to access system memory directly without CPU intervention, further reducing overhead.
• Sequential IO: Reading and writing data in a contiguous manner, optimizing disk access and throughput.

Data Transfer Without Zero Copy:

1. Disk to OS Cache: Data is initially loaded from the disk into the operating system's cache.
2. OS Cache to Kafka Application: The data is then copied from the OS cache into the Kafka application's memory space.
3. Kafka Application to Socket Buffer: Next, the data is copied from the Kafka application's memory to the socket buffer.
4. Socket Buffer to NIC Buffer: Finally, the data is copied from the socket buffer to the network interface card's buffer before being sent over the network.

This process involves four separate data copies and two system calls, making it inefficient.

Data Transfer With Zero Copy:

1. Disk to OS Cache: The initial step remains the same: data is loaded from the disk into the OS cache.
2. OS Cache to NIC Buffer (via sendfile): The Kafka application uses the sendfile system call to instruct the operating system to directly copy the data from the OS cache to the network interface card's buffer.

In this optimized path, there is only one data copy: from the OS cache to the NIC buffer.

DMA and Efficiency:

Modern network cards utilize Direct Memory Access (DMA). When DMA is employed, the CPU is not involved in the data transfer from the OS cache to the NIC buffer. This further enhances efficiency by offloading the copy operation to dedicated hardware.

Conclusion:

Kafka's high performance is attributed to two key design choices: sequential I/O and the zero-copy principle. By minimizing data copies and leveraging DMA, Kafka optimizes data transfer between disk and network, resulting in improved throughput and reduced latency. Kafka also uses other techniques to further optimize performance.