Understanding Zero‑Copy Techniques: DMA, sendfile, mmap and Direct I/O

Zero‑copy is a technique where the CPU does not have to copy data between memory regions; instead, data moves directly between devices or kernel buffers, reducing CPU work and context switches.

Four copies and four context switches occur in a naïve read‑write flow: from disk to page cache, page cache to socket buffer, kernel to user buffer, and user buffer back to kernel, each causing a user‑kernel transition.

Using DMA , a dedicated controller handles the data transfer between memory and I/O devices, allowing the CPU to issue only control commands. This eliminates two of the four copies and reduces context switches.

Zero‑copy implementations :

sendfile combines a read and write into a single system call, uses DMA and passes file descriptors to avoid copying data between kernel buffers.

mmap maps a file directly into user space, letting the application read/write the same pages without extra copies.

Direct I/O bypasses the page cache entirely, transferring data straight between user buffers and the device via DMA; it requires page pinning and may still need fsync for metadata.

Each method has trade‑offs: sendfile cannot be used when the data must be processed, mmap requires careful management of mapped regions, and Direct I/O adds complexity due to page pinning and lack of caching.

Typical use cases include high‑throughput message queues like Kafka, where producers write logs to disk using mmap and consumers send data over the network with sendfile, and databases that employ Direct I/O for self‑caching workloads.

The article concludes that Linux offers multiple zero‑copy strategies—reducing user‑kernel copies, bypassing the kernel, or optimizing buffer transfers—each suited to different performance requirements.