zerocopy

Zero-copy refers to techniques used to reduce or eliminate data copying during input/output operations, with the aim of lowering CPU overhead and memory bandwidth usage. The core idea is to move data between producers and consumers without creating intermediate buffers in user space or repeatedly duplicating memory contents. Zero-copy can occur within the kernel, or across user-space and kernel-space boundaries, and is a key performance tactic in high-throughput networking, file transfer, and I/O-intensive applications.

In network I/O, zero-copy techniques try to transfer data from a file or kernel buffer directly to a network interface, avoiding copies into user-space buffers. Classic examples include the Linux sendfile system call, which streams data from a file descriptor to a socket, and the use of splice or vmsplice to move kernel buffers without copying into user space. Some systems also employ kernel-bypass paths (for example, DPDK or SPDK) to let applications access NICs directly, further reducing copies. These approaches can reduce overhead, but may involve trade-offs such as complexity or partial copying for headers and control data.

In other contexts, zero-copy can involve memory mapping (mmap) to share pages between kernel and user space, thus avoiding explicit data copies, or using unbuffered/direct I/O to bypass kernel page caches. Modern I/O interfaces like io_uring on Linux offer features intended to minimize user-space copying by using registered buffers and asynchronous submission/completion paths. Nonetheless, true zero-copy depends on support across the hardware, kernel, and user-space libraries, and not all data transfers or platforms realize zero-copy benefits.

Limitations include portability concerns, the need for memory pinning or special buffer lifetimes, alignment requirements, and potential complexity in programming and maintenance.