SunOS 4.1 Performance Tuning

Hal L. Stern
Sun Microsystems
Northeast Area Consulting Group
Lincoln, MA
[email protected]

SunOS 4.1 includes a number of features designed to improve performance on both a per-application and system-wide basis. Features such as the tmpfs swap-based filesystem, NFS dynamic retransmission, mmap() and the associated madvise() hinting algorithm and dynamic allocation of kernel data structures may degrade performance if ignored or misconfigured. This paper presents performance analysis procedures and suggested policies for optimizing use of system facilities under normal conditions and avoiding severe performance degradation under peak work loads.

1. Introduction

Optimizing computing system performance requires efforts by both users and system administrators. Users expect system services to be well conditioned and to gracefully handle periods of peak demand; system managers rely on the user community to make efficient use of local and networked resources. Tuning the SunOS operating system does not require modifying the kernel source code, but instead entails establishing guidelines for system configuration and application development.

1.1. Performance Tuning in the Distributed Environment

Distributed computing systems add to the complexity of performance tuning. The user?s perception of system response time is dependent upon the composition of each of the system?s components: The code being executed, the local machine?s configuration, the network, and the performance of central servers.

System managers adopt a global view: performance tuning is geared toward meeting the aggregate demands placed on networks and servers. The global goal is to ensure acceptable response time during peak usage periods. Procedures used to achieve this goal involve varying configurations on both server and client machines. Increasing the size of kernel data structures on a server may provide part of a solution, while changing the NFS mounting schemes on the client side complements this process.

Users have a more local view of the tuning process. From a desktop workstation or terminal, the network and servers are generally of little interest to the user, particularly if