Text-only	Table of Contents (frame/ no frame)
(13) Performance Tuning

Performance Tuning

Distributed fileservers

AFS fileservers can be distributed throughout the network, so that in an inhomogeneous network environment, it is possible to have one or more servers more local in a network sense. Arranging to have the most frequently accessed volumes from a particular client on a close fileserver will improve performance.

Replicated Volumes

Volumes which are frequently read but only occasionally modified can be replicated on multiple servers. Clients will automatically pick a read-only copy if one is available, and will fail-over to a different copy if a server becomes unavailable. The top level volumes are usually replicated on every file server in a cell.

Replicating the master (read/write) copy onto the read-only copies is known as releasing the volume and is carried out manually, using the release subcommand to vos (vos_release)

Large vs Small Volumes

Since volumes are the unit of AFS server space for transactions such as migration between servers, replication, and backups, many operations are easier if the size is kept smaller. The important parameter is the ratio of the volume size to fileserver partition size. The daily backup snapshots can cause the actual disk usage of a volume to be double the visible space, if the files are modified daily. Large individual files make this problem worse. Backup snapshots limit the fraction of a server that can be allowed to fill. The tape backup system also requires staging space to hold the compressed backup snaphots, so large, frequently updated volumes have a significant impact.

Smaller volumes: pro

Easier to back up and restore
Easier to migrate between servers for load balancing or maintenance
More I/O bandwidth available, if the volumes are spread across multiple servers

Smaller volumes: con

Moving files between volumes involves copying data. Moving files between directories in the same volume is just changing a directory entry.
Simpler (slightly) permission management. ACLs on a volume top level are very important. It cannot be assumed to be protected by the ACL of a parent directory.
Simpler quota management.

Client cache options

AFS performance is very dependant on the speed and configuration of the client. Upgrading client CPU, disk, memory and networking all can have a large effect. The disk cache used on the client should if possible be on a disk partition of its own, ideally on a striped local filesystem. The client cache manager has various startup options which control how much memory it uses, and how much metadata related to volumes, directory entries and file data are kept cached locally. Increasing those values generally improves performance, but there is a point at which performance suffers because of searching times. The amount of file data transferred in one transaction (chunk size) can be tuned. Well connected fast networks with few lost packets will benefit from larger chunk sizes than the default. Graphical file managers will benefit from larger amounts of file and directory metadata in cache.

tuning.src last modified Jun 6, 2006

Introduction

Table of Contents
(frame/no frame)

Printable
(single file)