Robert Greig wrote:
On 17/10/06, lneves <[EMAIL PROTECTED]> wrote:

I've adopted Asyncweb (a Mina base HTTP server) do use your thread pool and tried to run some benchmarks against it, although it works, when I fire the load tester I get
lots of errors.

Read errors:

ERROR [SocketAcceptorIoProcessor-0.0] 17 Oct 2006 17:02:24 - Exception event
constructed: java.io.IOException: Connection reset by peer

What is on the other end of the connection? Is it not indicating an
error and as a result closing the connection?

On the other end I'm using the "Microsoft Application Center Test".
From what I can see MS-ACT is just simulating a bunch of users, creating and tearing down connections.

ERROR [SocketAcceptorIoProcessor-0.0] 17 Oct 2006 18:12:13 - Exception event
constructed: java.io.IOException: Broken pipe

Again it looks like the connection is being closed. Is there anything
in the logs on the other end of the connection?

The write errors occur at the end of the test where MS-ACT starts to close all connections.


These errors don't show up with the default thread pool.
Probably due to this errors I didn't get any throughput increase ... however as a
quick experiment I commented out the:
session.getFilterChain().fireExceptionCaught( session, e );
lines in the SocketIoProcessor and I did get a slight throughput increase with lower
CPU utilization.

We have done a lot of stress testing with our message broker and we
have never seen those errors. I am certainly very interested in
investigating further.


It could be that there is nothing wrong with your thread pool and this errors are being caused by some kind of Asyncweb misbehavior.... however, they don't show up with the default thread pool


What hardware are you running on? We are running all our tests with
two HP DL 585 boxes (4 dual core Opteron CPUs) connected with Gigabit
ethernet.

Nothing as muscular as that :-) ... the server is a single cpu machine running CentOS 4.4.

I have another server based on Mina (non HTTP) I will try to run some tests with it and report the results.

--
Luis Neves

Reply via email to