英语翻译We have evaluated the performance of a large e-commerce

英语翻译
We have evaluated the performance of a large e-commerce application through performance testing,performance monitoring and simulation modeling.We identified significant performance issues during both the performance testing and performance monitoring efforts.We used simulation modeling to help diagnose the cause of a very infrequent performance slowdown that was detected in the production environment and defied iagnosis for many months.The simulations we ran illuminated the fact that the variability introduced by the large 3 gigabyte heap size in combination with the kernel overhead induced by a large number of threads,caused the disruptive slowdowns reported by users.
We observed that good performance could be delivered by using any one of three architectural artifacts:vertical scalability,a 1 gigabyte heap size,or the use of a quality of service enforcement algorithm.The use of a 1 gigabyte heap has the drawback of requiring more frequent full garbage collection events but has the advantage of reducing the length of time for the garbage collection event.It provides a simple way of improving performance.
The use of vertical scalability benefits from load balancing and from better use of the available CPUs,since while one JVM may be performing a garbage collection,and single threading,the other JVMs may still make good use of the available CPUs.
The quality of service enforcement algorithm is designed to deny service to threads when the system is deemed too busy.It provides good response time at a cost of service
denial at heavy loads.If the application server has been properly tuned,we expect the quality of service enforcement algorithm to be activated only in emergencies,when the system is being subjected to a very heavy load,such as would be seen during a denial of service attack.
The most interesting result from our simulation modeling effort was that in spite of extensive performance modeling,measurement,and testing,we only gained insight into the system dynamics after the simulation model was built and its results analyzed.This was the case because the event that caused the slowdown only occurred when two infrequent events took place simultaneously,and it was very difficult to obtain data on the cause of the large backlog that occurred very infrequently in production.
In contrast,the model was able to quickly simulate a very large number of transactions.We are therefore convinced that performance studies could be efficiently run after the key factors that might affect the slowdown have been identified.