The performance of JSP pages is very close to that of servlets. However, users may experience a perceptible delay when a JSP page is accessed for the very first time. This is because the JSP page undergoes a "translation phase" wherein it is converted into a servlet by the JSP engine. Once this servlet is dynamically compiled and loaded into memory, it follows the servlet life cycle for request processing. Here, the jspInit() method is automatically invoked by the JSP engine upon loading the servlet, followed by the _jspService() method, which is responsible for request processing and replying to the client. Do note that the lifetime of this servlet is non-deterministic - it may be removed from memory at any time by the JSP engine for resource-related reasons. When this happens, the JSP engine automatically invokes the jspDestroy() method allowing the servlet to free any previously allocated resources.
Subsequent client requests to the JSP page do not result in a repeat of the translation phase as long as the servlet is cached in memory, and are directly handled by the servlet's service() method in a concurrent fashion (i.e. the service() method handles each client request within a seperate thread concurrently.)
There have been some recent studies contrasting the performance of servlets with Perl scripts running in a "real-life" environment. The results are favorable to servlets, especially when they are running in a clustered environment. For details, see:
http://www.objexcel.com/workingjava.htm#Web Server Benchmarks