ZeroCopy: Techniques, Benefits and Pitfalls

Abstract: We discuss various approaches intended to improve the data processing performance in OS kernels/drivers commonly described as ZeroCopy techniques. The main principle of ZeroCopy is to avoid completely or at least minimize unnecessary data copy operations by the CPU while processing I/O data in kernel drivers such as networking stacks and disk storage drivers. Modern CPU and memory architectures provide many interesting features to implement such techniques efficiently, however their advantages are often overestimated. The complexity of the memory architecture as well as necessary data processing in networking protocols may impose many problems and sometimes nullify the advantages of a promising ZeroCopy technique. We will try to take a closer look at various existing and proposed implementation and analyze their weak points. This paper will mostly concentrate on the modern UNIX-like operating systems such as Solaris and Linux with an emphasis on ZeroCopy performance and implementation issues in networking stacks.