On Web Server Performance Modeling for Accurate Simulation of End User Response Time

Transcription

1 On Web Server Performance Modeling for Accurate Simulation of End User Response Time Hiroshi Mineno NTT Service Integration Laboratories, NTT Corporation

2 Introduction Web systems Have rapidly become critically important IT systems Over 38 million websites ( The structure depends on where the logical structures are physically implemented. Networks Complicated by a mixture of applications, protocols, device and link technologies, traffic flows, and routing algorithms OPNET is very helpful to predict the performance of such complicated systems. But Are you really confident of your node parameter settings? No 2

3 Many reference materials Online documentation Methodologies & case studies SMARTE (Simulation Methodology for Application Response Time Engineering) Sample projects Past OPNETWORK proceedings But Are you really confident of your node parameter settings? This presentation will show: How to model the web server performance How to measure and set the parameters No 3

4 For predicting response time Client site IP IP cloud Server site End user response time Client Service Time Server Service Time Web browser Network delay for downloading SYN SYN+ACK ACK FIN ACK FIN ACK Web server 4 time Network delay for uploading Important modeling: HTTP/. (multiple persistent TCP connections) Network delay in each direction Service time at the client Service time at the server

5 Important modeling HTTP transaction Network delay in each direction Service time at the client Service time at the server Service time computation in a gna_clsvr_mgr.pr.m Request or Response size Service time = ( Service overhead + ) Processor load Processing speed ST = ST n Load n STn: average service time when n requests are concurrently processed at the server ST: basic performance, which is the average service time per request when one request is processed at the server Loadn: multi-tasking performance function, which degrades the basic performance of the server 5

6 OPNET actual setting ST /Load Loadn 6

7 ST: basic performance modeling Web server s function: Receive HTTP requests, produce response contents, then send them Types of content: - Static: html, jpg, gif - Dynamic: SSI (.shtml,.php,.asp ), CGI, NSAPI, ISAPI, FastCGI, Adapting ST calculations according to the content type: ST ST ST static = FS PS FS static dynamic dynamic = + PS OH Processing time depends on the content size Remaining service overhead -STstatic: ST for static content - STdynamic: ST for dynamically generated content -FSstatic: size of static content - FSdynamic: size of dynamically generated content - PS: processing speed at application layer - OH: remaining service overhead 7

8 Loadn: multi-tasking performance modeling Web browser RT SYN SYN+ACK ACK HTTP: Get ~.html HTTP: Get ~.gif HTTP: Get ~.jpg Web server ST Web browser RTn SYN SYN+ACK ACK HTTP: Get ~.html HTTP: Get ~.gif HTTP: Get ~.jpg Web server STn One request is processed at the server. Load n = ST ST RT RT 8 n N requests are concurrently processed. n

9 How to measure the parameters for STn Capture the packet trace for ST Use one virtual user Virtual users Protocol analyzer (Sniffer Pro) Number of virtual users 48 Short cross cable Stress test tool (LoadRunner) Web server 3 0 Elapsed time of stress test Measure the response time for Loadn 9

10 Setting the parameters of ST Web browser Web server Analyzing captured data exchanges MST MST2 MSTp- MSTp - p: total number of contents - q: number of dynamic contents - MST: measured service time at the server 0 p q FS i= 0 i MSTi = OH j + i= 0 j= 0 PS FS k MSTk = OH k + ( k = 0, K, q) PS ST parameters = p i FSi 0 FS dynamic p i = 0 MST dynamic MSTi p Webpage-A p 8 q 54,007 bytes 322 bytes 27.4 ms ms PS 7,727,796 bytes/s OH ms OH

11 Validating ST modeling Experimental environment Simulation model Short cross cable Client Win2000 Pro LoadRunner 7 Sniffer Pro Webpage-A Measured Simulated Web server Win2000 server IIS5 Response time (average) ms ms Modified server node: -Adapt ST calculations -Set the estimated PS, OH

12 Validating data exchanges of ST modeling Measured data exchanges ms (average) Simulated data exchanges ms (average) 2

13 Validating AppDoctor statistics of ST modeling Measured statistics Simulated statistics 3

14 How to measure RTn Repeat requests immediately to keep number of active users constant Increase from one virtual user per 30 seconds up to 48 Pay attention to three points:. Before measuring the data for RTn, make some virtual users repeatedly hit the cache 2. Each virtual user should use only one TCP connection Average response time per second (sec) 4 Number of virtual users Elapsed time of stress test 0:00 4:48 9:36 4:24 9:2 24:00 28:48 Elapsed time of stress test (min:sec) Average response time per second for all the active users.

15 Setting the parameters of Loadn 3. Remove delays caused by the network and the client ExtraDelay= RT Load n Multi-tasking performance table = RT RT n p i = o MST ExtraDelay ExtraDelay /Load Loadn i y = x R 2 = Number of virtual users RT Loadn parameters RT 5 = p i Average response time of each virtual user and /Loadn. 0 MSTi Webpage-A bytes ms ExtraDelay RTn ms Average response time (s)

16 Validating Loadn modeling Experimental environment Simulation model Short cross cable Client Win2000 Pro LoadRunner 7 Number of virtual users Web server Win2000 Server IIS5 Elapsed time of stress test Modified server node: -Adapt ST calculations -Set the estimated PS, OH (when web server s cache was hit) -Multi-tasking performance table: /Load Loadn, /X Individually repeat requests in a uniformly distributed time cycle after getting their own response 6

17 Results of Loadn modeling Average response time (s) Stress test Simulated -Loadn- Simulated -default Request sending rate (requests/s) 7

18 Comparison with using ACE ACE allows us to model an application s behavior accurately as well as easily. We investigated how ACE can allow us to model STn. 8

19 Validating data exchanges of ST modeling with ACE Measured data exchanges ms (average) Data exchanges simulated by using ACE file ms (average) 9

20 Validating AppDoctor statistics of ST modeling with ACE Measured statistics Statistics simulated by using ACE file 20

21 Results of STn modeling with ACE Average response time (s) Stress test Simulated -Loadn- Simulated -default- Simulated -Imported by using ACE Request sending rate (requests/s) Unable to apply the performance degradation Loadn!! 2

22 Conclusion This presentation showed: How to model the web server performance STn How to measure and set the parameters for STn Future work: Now Are you confident of your node parameter settings? Yes! Apply Loadn to the server performance even if when ST was imported by ACE Develop the node performance measurement module 22