7 Responses

1. Trustin Lee

   March 4th, 2009 at 9:38 am

   1

   Hi Mike, thanks for sharing your test result! The number looks pretty nice.

   To correct just one thing, any write request in Netty is supposed to be queued to its internal buffer and flushed by I/O worker threads. This is pretty same with what MINA does.

   Netty worked just like you mentioned in your post before, but the behavior has been reverted back when 3.0.0.GA is released due to some contention. I’d like to figure out the root cause of the contention to reduce the latency even more soon.

   I believe where the biggest difference comes now from is probably the effective usage of data structure. Especially, Netty has much more effective buffer and queue implementation.

2. Trustin Lee

   March 4th, 2009 at 9:51 am

   2

   I realized that they are not exactly same at handling write requests. If the write() is called from an I/O thread, Netty tries to flush the buffer as soon as possible, while MINA doesn’t flush the queue to a certain point.

   If the write() is called from an other thread than an I/O thread, both behaves same. However, that’s where the better queue and thread pool implementation kicks in.

3. emmanuel lecharny

   March 5th, 2009 at 11:57 am

   3

   Hi Mike,

   what would be interesting is to profile MINA to know exactly where the time is lost. I’m not sure that the bottleneck is where you think… Or if it’s there, then, well, you are a MINA committer, isn’t it ?

   What’s obvious is that we have many things to improve in MINA. Btw, the number you gave just shiw a slight difference, around 10%, which is still significant, but not overkilling. Except if you consider that a large part of the time is consummed in MySQL.

   Last, not least, the very same test without any application at the end - ie, just a plain loopback - might demonstrate a bigger difference. Could you strip your test from the MySQL part, and re-run it ?

   Would it be too much to ask you to share the test environment too ?

   Many thanks !

4. Niall

   March 28th, 2009 at 6:12 pm

   4

   From your results thats about 10% faster - I guess we have different definitions of significant

5. Scalable NIO Servers - Part I - Performance « Z|NET Development

   April 7th, 2009 at 8:55 pm

   5

   [...] http://swamp.homelinux.net/blog/index.php/2009/03/03/performance-comparison-of-apache-mina-and-jboss... [...]

6. jim

   May 28th, 2009 at 7:23 pm

   6

   my project 1xio at Google Code http://code.google.com/p/1xio takes a non-managed approach to NIO async http stream serving.

   I implemented Selectors for write which is uncommon but it works out nicely in the single thread environment.

   in this implementation I use NIO as an intentional java memory leak for it’s NIO objects and do what i can to avoid strings by marking up the input ByteBuffers rather than manufacture new heap objects.

   I added mimetypes and the UTF8 java strings roughly equivalent to /etc/mimetype in size double the footprint.

   there is a single thread, 4 enums, and no classes presently to provide a reasonably useful static http server.

   i use a technique I call ‘java metaprogramming’ similar in the spirit of c++ template metaprogramming which might have some in-constructor jit benefits over other static singleton visitor pattern strategies for asynchronous services.

7.   Performance comparison of Apache MINA and JBoss Netty Revisited by Totally Nerdy Information

   November 27th, 2009 at 12:25 pm

   7

   [...] I ran my my original comparison, both the tests and databases were running on the same host connecting over the loopback interface. [...]