Performance Modeling and Benchmarking of Event-Based ... - DVS
Performance Modeling and Benchmarking of Event-Based ... - DVS
Performance Modeling and Benchmarking of Event-Based ... - DVS
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
7.1. ONGOING AND FUTURE WORK 137<br />
JA(27%-+09+,#5"76#+<br />
D0%(>0$)E+F11276(>0$)+($*++<br />
G$*#52-7$8+H#6:$02087#)+<br />
@A$>.#+B0$7%057$8+C+<br />
45#*76>0$+<br />
!"#$%&'()#*+,-)%#.)+<br />
/0.12#3+!"#$%+4506#))7$8+<br />
0$+++<br />
'#$6:.(5;7$8+09+!',++<br />
'#$6:.(5;7$8+<br />
B#%:0*0208-+<br />
4#5905.($6#++<br />
!$87$##57$8+<br />
'#$6:.(5;+<br />
!$87$##57$8+<br />
D#I+H5#$*)+<br />
!"#$%&'!$$%()"(!$$%&'*'<br />
Figure 7.1: Open Research Issues<br />
first case study we extended SIENA, a well-known DEBS, with a runtime measurement framework.<br />
We evaluated a system with 15 brokers, 8 publishers <strong>and</strong> 16 subscribers <strong>and</strong> predicted the<br />
runtime behavior including delivery latency for a basic workload with a single event type. In the<br />
second case study, we developed a comprehensive model <strong>of</strong> the complete SPECjms2007 workload<br />
including the persistent layer, point-to-point <strong>and</strong> publish/subscribe communication. We applied<br />
several <strong>of</strong> our performance modeling patterns. Furthermore, the workload was modeled using<br />
the proposed QPN extensions.<br />
We evaluated its accuracy in a commercial middleware environment. To validate our modeling<br />
technique we investigated deployments <strong>of</strong> the benchmark in representative environments<br />
comparing the model predictions against measurements on the real systems. A number <strong>of</strong> different<br />
scenarios with varying workload intensity (up to 30,000 messages / 4,500 transaction per<br />
second) <strong>and</strong> interaction mixes were taken into account. By means <strong>of</strong> the proposed models we<br />
were able to predict the performance accurately. To the best <strong>of</strong> our knowledge, no models <strong>of</strong><br />
realistic systems <strong>of</strong> the size <strong>and</strong> complexity <strong>of</strong> the one considered in this thesis exist in the<br />
literature.<br />
The results <strong>of</strong> both case studies demonstrated the effectiveness <strong>and</strong> practicality <strong>of</strong> the proposed<br />
modeling <strong>and</strong> prediction methodology in the context <strong>of</strong> a real-world scenario. The advantage<br />
<strong>of</strong> the proposed approach is that it is both practical <strong>and</strong> general, <strong>and</strong> it can be readily<br />
applied for performance evaluation <strong>of</strong> DEBS <strong>and</strong> MOM. The technique can be exploited as a<br />
powerful tool for performance prediction <strong>and</strong> capacity planning during the s<strong>of</strong>tware engineering<br />
lifecycle <strong>of</strong> message-oriented event-driven systems.<br />
7.1 Ongoing <strong>and</strong> Future Work<br />
<strong>Based</strong> on our experience <strong>and</strong> the findings <strong>of</strong> this work, we identified three interesting areas for<br />
future research:<br />
• Benchmark engineering<br />
• Workload characterization <strong>and</strong> benchmarking <strong>of</strong> event-based systems