Posted by driscoll
on June 28, 2005 at 4:02 PM PDT
In the Grizzly tech session, the performance team from Sun talked briefly about some Grizzly (NIO) performance numbers.
I'm in the Grizzly Tech session, and at the end, they're showing some performance numbers.
For a dynamic content, servlet oriented, non-trivial task, the numbers that they've found show that an all Java NIO implementation will beat out a mixed Java-C implementation. (They tested two.) Mixed implementations are the kind found today in things like mod_jk and Sun's Web Server (though they wouldn't say which they tested). Grizzly's the first NIO type implementation that we know of out there, but if these numbers hold up (and we've beaten on 'em), it won't be the last.
Why is NIO better? Crossing the JNI boundary is more expensive than the small amount that you'd give up choosing NIO over C.
Interesting, eh? (Well, it is to me.)
Oh, and of course both mixed-C and NIO totally blew away a traditional implementation, since you're thread bound and limited by the OS. But that part shouldn't suprise anybody.
Got benchmarks that say different? I'd like to hear about them...