Reducing interprocess communication overhead through register windows

Inter-process communication (IPC) by message passing is a common paradigm of many operating system architectures. IPC performance is a critical part of overall system performance. Unfortunately, most architectures exhibit extremely high overhead for short communication. We present a combination of hardware and software primitives that effectively eliminate overhead in short message transfers. CPU registers act as a natural interface between task address spaces, making IPC more efficient and easier to implement than existing software solutions. Our performance study shows that using dynamically relocatable register windows can improve overall system performance by more than fifty percent over typical modern system architectures