On 09/05/2022 19:01, Cay Horstmann wrote:
Il 03/05/2022 08:31, Cay Horstmann ha scritto:
Il 03/05/2022 07:25, Rémy Maucherat ha scritto:
On Tue, May 3, 2022 at 6:24 AM Cay Horstmann
<cay.horstm...@gmail.com> wrote:
Hi, I am trying to experiment with Loom
(https://openjdk.java.net/jeps/425) virtual threads in Tomcat 10. There
is a nice extension point in server.xml where I can provide an Executor
and use it in the default Connector. It works like a charm--all my
requests run on a virtual thread, and AFAIK, nothing else does.
There is just one issue. It takes forever for Tomcat to execute the
virtual threads. I made a timestamp for each invocation of
public void execute(Runnable command)
called from org.apache.tomcat.util.net.AbstractEndpoint.processSocket
When I fire 1000 simultaneous requests, I can see that method being
invoked 1000 times, but it takes a while to work through the requests.
The last invocation occurs two minutes (!) from the sending of the
requests.
When using the org.apache.catalina.core.StandardThreadExecutor, that
many requests are handled in 3 seconds.
Is there something that I am overlooking? I had hoped that execute
would
be called near-instantly 1000 times, and then the Loom virtual threads
could show their mettle and execute concurrently.
I wanted to experiment with Loom, but most likely the current Tomcat
NIO(2) connector is not nice for that. I was thinking that
resurrecting the java.io code (actually: writing a new java.io
connector) could be a slightly better plan, but to be honest I don't
expect very good results. I wasn't planning to do it immediately since
Loom is so experimental right now.
I figured this out in the end. When using Nio2, changing the executor in
server.xml doesn't seem useful. That executor is then used in the inner
workings of the selector mechanism. That is probably not what anyone
wants to do. But with Nio, plugging in Loom for running the servlet
methods is very reasonable. There is only one issue. Loom cannot (yet)
handle "synchronized".
Changing the synchronized block in
org/apache/tomcat/util/net/SocketProcessorBase.run() to a ReentrantLock
makes the Loom threads fly, with impressive performance.
Care to share the code you used to test this? I'd like to see if I can
reproduce the results and get some numbers on the performance difference.
Thanks,
Mark
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@tomcat.apache.org
For additional commands, e-mail: dev-h...@tomcat.apache.org
