Tapestry作者对Tapestry和JSP性能比较的看法

It's accepted knowledge that JSPs are efficient because they are monolithic, stateless, compiled Java classes.

However, like much accepted, common knowledge, it may not be correct.

I've had to debug some problems in JSPs using Struts in both Tomcat and WebLogic, so I've seen the kind of code that gets generated.

What I found interesting is the amount of code related to pooling JSP tag instances. The thing is, the JSP as a whole is stateless, but while the JSP is rendering, the tags within the JSP do have considerable transient state.

So, each time a JSP tag is referenced in a JSP, the corresponding code obtains a pooled instance of the tag, re-initializes all its attributes, uses it ... then clears it and puts it right back into the pool.

This is where Tapestry has a bit of an advantage; in Tapestry, the entire *page* is pooled, not the components within the page, so the cost of a component is relatively small. Further, the way Tapestry parameters work allows a component to identify intransient parameters ... parameters whose value doesn't change, such as with a <string-binding> or <message-binding>, or a literal value in an HTML template. In those cases, the component doesn't need to reset the property value after the component renders, or re-initialize the property value when the component is utilized again.

Tapestry 3.1 will be using even smarter dynamic bytecode generation (using Javassist) to further improve the efficiency with which a component can access its parameters. OGNL will be using Javassist (or a similar library) to build more efficient access paths to object properties, as efficient as ordinary method invocations, rather than reflection.

So, the optimizations (including the invariant checks), mean that Tapestry is using OGNL less ... and the OGNL improvements mean that when OGNL is used, it works more efficiently.

Now all we have to do is build "equivalent" applications and start comparing.

That, combined with Drew's upcoming improvements to OGNL should be very exciting. Mind Bridge has profiled Tapestry applications and they can spend around 50% of their time evaluating OGNL expressions. The majority of those OGNL expressions are simple property references. They're working on a plan to make use of Javassist to replace reflection with direct method invocations where appropriate, and Drew's seen about a 10x improvement.