Bill Barker wrote:
So, where is it already? Some of use want to see how you can beat
Costin ;-).
Right now, I'm not planning on beating anyone, it's just my crap work
code (which doesn't actually work yet). I haven't benched anything, and
my experience with AJP is very limited.
I used the package org.apache.coyote.ajp for it, and it reuses most of
the message code, parsing and composition from regular AJP. One of the
issues I'm facing, however, is that the usage of JNI will make the
current code subpar, where it would be a lot better to do message
reading and writing (this means composition of the messages to avoid
useless bytes copying) using direct buffers. So there's a lot of work
left, and it will remain experimental for a little while.
If I understand you correctly, you want MsgAjp to use ByteBuffer instead
of byte []. At the cost of never supporting JDK 1.3 ever again, this
would probably actually improve the performance of ChannelSocket (after
changing it to use a blocking SocketChannel).
Well, yes, I'll have to use a direct ByteBuffer like I do for HTTP.
Otherwise, byte array copy over JNI will be slower. It's quite annoying,
but using JNI imposes a certain design for buffering :(
One of the possibilities is to have AjpMessage (I renamed it, sorry)
backed by a direct BB, or - as I am investigating at the moment - just
write directly to a large BB. Reading, which faces the same problem,
will likely have to use a large direct BB to avoid useless copying. For
example, the current AJP code does two reads to read a message. Since in
JNI, the best optimization is to lower the amount of JNI calls, this
will have to be changed (which has fairly large consequences).
Rémy
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
