Performance Assessments of Internet Systems By Kishore G

Performance Assessments of Internet Systems By Kishore G. Kamath SPAN Technologies Testing solutions for the enterprise www. spantechnologies. com Kishore@spantechnologies. com 973 -575 -7235 ã Copyright 2003 by SPAN Technologies.

Agenda • • • Why do Internet systems under perform? Performance analysis in the corporate world How do automated tools work? Simulating user load Assessing computer systems ã Copyright 2003 by SPAN Technologies.

Performance Considerations for Internet Applications • Why do Internet applications under-perform? – Application Design/Implementation » In-efficient database queries » Poorly written code • Identification Web Users – Simulate light user-load – Measure system metrics on Database / App Server » CPU Utilization, Disk I/O and Processor Queue Length Web Server Database Server App Server ã Copyright 2003 by SPAN Technologies.

Performance Considerations for Internet Applications Courtesy Mercury Interactive Corp ã Copyright 2003 by SPAN

Performance Considerations for Internet Applications • Why do Internet applications underperform? – Unbalanced System Topology » Physical design and topology may be in -efficient » App Server components may not be colocated on same box. • Identification – Characteristics » High network load » Moderate CPU Utilization ã Copyright 2003 by SPAN Technologies. Web Server 2 Load balancer – Simulate moderate user-load – Measure system metrics on Database / App Server » CPU Utilization, Disk I/O and Processor Queue Length • Measure network metrics Database server App server

Performance Considerations for Internet Applications • Why do Internet applications under-perform? – In-adequate Server/Network Capacity » Server hardware under equipped to handle software demands. » Network pipe to the Internet is too small. • Identification – Simulate anticipated user-load – Measure system metrics on Database / App Server » CPU Utilization, Disk I/O and Processor Queue Length – Measure network metrics – Characteristics » High network load / High CPU Utilization, Disk I/O or Processor Queue Length. ã Copyright 2003 by SPAN Technologies.

Performance analysis in the corporate world • Create a Test Region a scale of production • Simulate user activity using an automated performance tool (Mercury Interactive Load. Runner, Compuware QALoad, Empirix E -Load or Segue Silk Performer) • Use automated tools to – Record user activity – Replay user activity x number of users to simulate. • Examine system behavior under simulated load – Review system metrics – Do root-cause analysis. ã Copyright 2003 by SPAN Technologies.

How do automated tools work? • How recording works – Capture user activity at a protocol level – Create a proxy server to capture HTTP requests from a browser – Capture HTTP request components including server target requested and name value pairs. – Store these HTTP requests into a script. Record Web server App server 1 Private HTTP Request Proxy Script name = “jane” password = “courant” ã Copyright 2003 by SPAN Technologies. Database server

How do automated tools work? • How replay and analysis work – Make HTTP requests stored in script. – Replace pivotal values in name-value pairs with other values. » Prevent caching from making results optimistic. Replay name = “jane” password = “courant” Web server Script name = “john” password =NYU” name = “harry” password = “edu” App server HTTP Request Script ã Copyright 2003 by SPAN Technologies. Database server

How do automated tools work? • How replay and analysis work – Server cannot tell apart a script replay from a real-user request. – Multiple instances of script execution can simulate multiple users. » Instances can simulate increasing, decreasing, steady-state or random user load. – HTTP requests and server response and behavior metrics can be monitored and collated. – Correlations between requests and behavior can be analyzed. ã Copyright 2003 by SPAN Technologies.

Performance analysis • Light load – Load is less than 10% of expected load. – Review of transaction response times and server/network metrics » High response time coupled with high Disk I/O or CPU Utilization is indicative of bad code/queries. ã Copyright 2003 by SPAN Technologies.

Performance analysis Courtesy Mercury Interactive Corp ã Copyright 2003 by SPAN Technologies.

Performance analysis • Steady state 100% load – Review of server/network metrics » Increasing trend indicates an issue » Memory leaks can be determined Courtesy Mercury Interactive Corp ã Copyright 2003 by SPAN Technologies.

Performance analysis • Capacity Planning – Increasing load past expected load levels » Review of response times and server/network metrics • Response time under expected and future load. • Server/network utilization under expected and future load Courtesy Mercury Interactive Corp ã Copyright 2003 by SPAN Technologies.

Comparative studies • Run load under 2 different conditions – Example » With and without SSL. » Two different server hardware configurations. » Two different configuration settings • Compare response times and metrics to study impact of change – Used for tuning, benchmarking. ã Copyright 2003 by SPAN Technologies.

Courtesy Mercury Interactive Corp ã Copyright 2003 by SPAN Technologies.

IP based load-balancing • Load re-directors will send load from one IP address • • to the same web server. Load simulation would need to use different IP addresses to simulate multiple users Load simulator will spoof IP Addresses not in use when making requests to the server. – Max of 25 -8192 IP addresses per Network card depending upon OS. • Load balancer uses IP address to direct load to • appropriate web-server based on hashing algorithm Server host alias file re-points IP addresses back to the true load simulator IP address. ã Copyright 2003 by SPAN Technologies.

Summary • Performance assessments of Internet systems • • are achieved via simulation. Automated tools can create realistic load by simulating HTTP requests while changing data. Correlation of system metrics with level of load can identify performance issues and their root cause. ã Copyright 2003 by SPAN Technologies.