Scattergather

Scatter-gather refers to a class of techniques for transferring data where the memory involved is non-contiguous. Instead of moving a single contiguous block, a list of address-length pairs is used to describe multiple chunks of memory. The operation can scatter data from a source to multiple destinations or gather data from multiple sources into a single destination. The list is commonly called a scatter-gather list or an I/O vector.

In parallel and distributed computing, scatter and gather are common collective operations. MPI, for example, provides Scatter to distribute elements from a root process to all processes and Gather to collect elements from all processes to the root. Variants such as Scatterv and Gatherv support variable counts per process, enabling flexible distribution and collection patterns for parallel workloads.

In I/O and hardware, scatter-gather appears in two related contexts. Scatter-gather I/O uses an array of buffers to perform a single logical I/O across multiple non-contiguous user buffers, typically via readv and writev system calls. In hardware and operating systems, scatter-gather DMA uses a scatter-gather list to move data between non-contiguous regions of memory and a device, reducing the need for intermediate copies and improving throughput.

Benefits of scatter-gather include reduced data copying, efficient handling of fragmented buffers, and better utilization of non-contiguous memory layouts. Its drawbacks involve added complexity in addressing, potential overhead in managing the scatter-gather lists, and platform-specific API differences. Common applications include network stacks, high-performance file and memory I/O, and accelerator or device DMA engines.