28.11.2014 Views

Performance Tuning Guide - EMC Community Network

Performance Tuning Guide - EMC Community Network

Performance Tuning Guide - EMC Community Network

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Measuring <strong>Performance</strong><br />

Analyzing multi-user tests<br />

After completing your multi-user tests, analyze the following areas for bottlenecks:<br />

• Determine which server or process consumed the most CPU or memory.<br />

• Check the database server for queries that take a long time, deadlock or core dump alerts, and<br />

top wait events. For Oracle user, use an AWR report for your analysis (See Assessing database<br />

bottlenecks and dead waits, page 84).<br />

• Check the J2EE layer for excessive garbage collections, which indicates a bottleneck in the<br />

application server. If so, add additional JVM instances to improve load balancing.<br />

• Check for bottlenecks indicated in the Content Server repository and BPM logs, and the<br />

application server logs.<br />

• Check for network bottlenecks indicated by bandwidth exhaustion or by exceeding the maximum<br />

number of packets per second that can be transferred. Even though network bandwidth is not<br />

exhausted, a typical network card can only handle definite number of packets a second. In this<br />

case adding additional network cards can help.<br />

When addressing a problem area, incrementally make one change at a time so that you can<br />

understand the impact of the change, then rerun the test. Depending on the result of the rerun<br />

test, you can either apply or reject your change.<br />

Avoiding multi-user test worst practices<br />

The following provides activities to avoid during multi-user testing:<br />

• Do not turn on excessive logging or tracing. The volume of log files generated in a multi-user<br />

load test makes analysis of results difficult.<br />

• Do not synchronize users to execute an operation at the same time. Interpretation of these types of<br />

results regarding the true capability of your system can be misleading.<br />

• Avoid simulating super-users. Using a few super-users to process requests is not the same as<br />

spreading the same number of requests over a larger set of normal-users. For the normal-user<br />

scenario, each user creates a unique backend session with caching and history being tracked under<br />

that particular session. The purging and reuse of cache data is based on the wall-clock request<br />

time of that user session. A super-user session violates most of the built-in design algorithms<br />

because request times are unrealistically condensed.<br />

• Do not exceed 80% CPU usage or memory capacity on your servers. During normal usage,<br />

operating systems require 20% excess capacity to handle naturally occurring usage spikes.<br />

Red-lining servers generates results that are misleading and bring no value to the overall<br />

validation of the system. Be prepared to add servers when 80% capacity is reached.<br />

• Avoid running multi-user testing on unstable builds. Make sure your single user testing is<br />

acceptable before doing multi-user testing.<br />

<strong>EMC</strong> Documentum xCP 1.0 <strong>Performance</strong> <strong>Tuning</strong> <strong>Guide</strong> 83

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!