DTJ508PF

This document describes a methodology for capacity planning in PATHWORKS client-server environments, which provide remote file services to desktop PCs.

The core challenge in these environments is that server resource utilization is highly variable, depending on client applications and user techniques, which can change significantly over time (e.g., through macro usage). While capacity planning aims to estimate hardware needs and explore "what-if" scenarios, accurately measuring client-perceived response time and the exact number of active users from the server side is difficult, making it an "open model."

The paper proposes using a queuing analytical model, primarily with DECperformance Solution software. The process involves:

1. Data Collection: Gathering detailed resource consumption data from the server, supplemented by subjective user performance evaluations, over typical workdays to capture real-world variation.
2. Workload Classification: Grouping server processes into distinct "workload classes" (e.g., FILESVS, OVERHEAD) for independent analysis and manipulation.
3. Model Building & Manipulation: Building a validated model, normalizing it by removing "abnormal" processes, and then adjusting user numbers or hardware configurations to analyze their impact on server response times (which serve as indicators for client-side performance).

Key findings from this analytical modeling include:

• Significant Impact of User Behavior: Even minor changes in user work habits (like running macros) can drastically increase server I/O and queuing, severely impacting performance and leading to bottlenecks.
• CPU Performance Paradox: When maintaining a uniform server response time, slower CPUs are found to be less utilized than faster ones. This is because their longer service times make them more susceptible to overhead, reducing available processing time for the primary workload.
• High Workload Variability: Day-to-day user activity and resource consumption on file servers can vary by a factor of 3 to 5, necessitating robust capacity planning to accommodate peak loads.

The analysis provides insights into the causes and symptoms of server resource exhaustion, enabling capacity planners to predict server response times and make informed decisions about system sizing and upgrades.