5.2 Alternative Network Technologies

		as though they were local files. Unlike other protocols, an NFS server is stateless, meaning it doesn't save any information about a client between requests. This means that all client requests are considered independently and must therefore contain all the information necessary for execution. All NFS read and write requests must include file offsets, unlike local file reads and writes which proceed from where the last one left off. The stateless nature of the NFS server causes messages to be larger, potentially consuming network bandwidth. The advantage of statelessness is that the server is not affected when a client crashes. The best way to configure NFS on a Beowulf system is to minimize the number of mount points, set the read and write buffer sizes to the maximum allowable values (8192 bytes), and use the autofs daemon discussed later in this section. You can set the buffer sizes using the rsize and wsize options for the NFS file systems listed /etc/fstab. A typical fstab entry for mounting /home may look like the following:

		b001:/home /home nfs rw,hard,intr,bg,rsize=8192,wsize=8192 0 0

		The original Linux NFS implementation only allowed a single NFS server to run at a time. This presented severe scaling problems for Beowulf clusters, where many internal nodes would mount home directories and other file systems from the worldly node. A single NFS server would serialize all network file system accesses, creating a severe bottleneck for disk writes. Disk reads were not as adversely impacted because the clients would cache files locally. More recent versions of the Linux NFS implementation allowed multiple servers operating in read-only mode. While this was useful for certain LAN applications, where workstations might mount readonly /usr/ partitions, it was not of such great benefit to Beowulf clusters, where internal nodes frequently require NFS for performing disk writes. The very latest versions of the Linux NFS code released in 1998, starting with version 2.2beta32, have added support for multiple servers in read/write mode. This new functionality has not yet been extensively tested, and the performance gains not fully quantified. However, it is expected that this addition will increase the scalability of NFS as used in Beowulf clusters.

		5.6.2 AFS

		The Andrew File System (AFS) was originally developed at Carnegie Mellon University as a joint venture with IBM in the mid-80s. Its purpose was to overcome the scaling problems of other network file systems such as NFS. AFS proved to be able to reduce CPU usage and network traffic while still delivering efficient file system access for larger numbers of clients. In 1989, development of AFS was transferred to Transarc Corporation, who evolved AFS into the Distributed File System