Table of Contents
Scalability for SMP Systems
Performance Defined
WHY Does Performance Matter?
Why Use a Multi-Processor System
Will SMP Improve Existing Apps?
How New Apps Can Exploit SMP
Scalability Defined (per W.R.S.)
Ideal Scalability Curve for MP
Limitations to Scalability
Steps in Evaluating Scalability
Common Measurements
Common Measurements (cont.)
Case Study An Online Web Based App.
Analysis and Tuning Methodology
Application Performance on ES6000 20way for 1000 requests - Elapsed time 850 seconds and 3500 CPU seconds with 10 simultaneous users
System Level CPU Usage
Application Thread Activity
Application Scheduling Delay
Scalability Projection
Scalability Projection (cont)
What Caused System Time?
Mutex Contention - How it occurs
Mutex Contention - What it Costs
Mutex Contention - What it Costs
Mutex Contention - Where it occurs
After Tuning the SQL queries in RW DBTools 1000 Requests take 600s and use 2525 CPU Seconds
Mutex Contention Still Persists after addition of RW Mutex Pool Enhancement 1000 Requests take 600s and use 2698 CPU seconds
Smart Heap Reduces Mutex Contention (without RW mpool) 1000 requests in 250s and 1165 CPU seconds
A Closer Look Reveals Periodic Mutex Contention at Points of Low Messaging Activity
No Contention After Adding RWT Mutex Pool Modification 1000 requests in 110s using 350 CPU seconds
Thread Activity
Scalability Projection
Current Production Status
Conclusions Drawn
Common Fallacies
A Case Study of an ERP App
Signs of Mutex Contention
Total System CPU
Number of Active Threads
Quantifying Scalability
Non-Linear CPU for User and System
Program is Limited by Response XMIT
Poor Communication Programming
Projected Improvement (XMIT only)
Projected Improvements Total
Design and Analysis Approach
Conclusion
|