Check out the new USENIX Web site. next up previous
Next: Performance Up: IO-Lite: A Unified I/O Previous: Operation in a Web


This section gives an overview of the implementation of the IO-Lite prototype system in a 4.4BSD derived UNIX kernel. IO-Lite is implemented as a loadable kernel module that can be dynamically linked to a slightly modified FreeBSD 2.2.6 kernel. A runtime library must be linked with applications wishing to use the IO-Lite API. This library provides the buffer aggregate manipulation routines and stubs for the IO-Lite system calls.

Network Subsystem: The BSD network subsystem was adapted by encapsulating IO-Lite buffers inside the BSD native buffer abstraction, mbufs. This approach avoids intrusive and widespread source code modifications.

The encapsulation was accomplished by using the mbuf out-of-line pointer to refer to an IO-Lite buffer, thus maintaining compatibility with the BSD network subsystem in a very simple, efficient manner. Small data items such as network packet headers are still stored inline in mbufs, but the performance critical bulk data resides in IO-Lite buffers. Since the mbuf data structure remains essentially unmodified, the bulk of the network subsystem (including all network protocols) works unmodified with mbuf encapsulated IO-Lite buffers.

Filesystem: The IO-Lite file cache module replaces the unified buffer cache module found in 4.4BSD derived systems [17]. The bulk of the filesystem code (below the block-oriented file read/write interface) remains unmodified. As in the original BSD kernel, the filesystem continues to use the ``old'' buffer cache to hold filesystem metadata.

The original UNIX read and write system calls for files are implemented by IO-Lite for backward compatibility; a data copy operation is used to move data between application buffers and IO-Lite buffers.

VM System: Adding IO-Lite does not require any significant changes to the BSD VM system [17]. IO-Lite uses standard interfaces exported by the VM system to create a VM object that represents the IO-Lite window, map that object into kernel and user process address spaces, and to provide page-in and page-out handlers for the IO-Lite buffers.

The page-in and page-out handlers use information maintained by the IO-Lite file cache module to determine the disk locations that provide backing store for a given IO-Lite buffer page. The replacement policy for IO-Lite buffers and the IO-Lite file cache is implemented by the page-out handler, in cooperation with the IO-Lite file cache module.

IPC System: The IO-Lite system adds a modified implementation of the BSD IPC facilities. This implementation is used whenever a process uses the IO-Lite read/write operations on a BSD pipe or Unix domain socket. If the processes on both ends of a pipe or Unix domain socket-pair use the IO-Lite API, then the data transfer proceeds copy-free by passing the associated IO-Lite buffers by reference. The IO-Lite system ensures that all pages occupied by these IO-Lite buffers are readable in the receiving domain, using standard VM operations.

next up previous
Next: Performance Up: IO-Lite: A Unified I/O Previous: Operation in a Web
Peter Druschel