mina-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Emmanuel LŽcharny <elecha...@gmail.com>
Subject Re: [3.0] IoBuffer
Date Sun, 13 Dec 2009 10:48:54 GMT
Bogdan Ciprian Pistol a écrit :
> Hello all,
>
> A discussion on the https://issues.apache.org/jira/browse/DIRMINA-751
> arose that is more appropriate here on the ML.
>
> Emmanuel Lecharny said:
> "Some thoughts :
> - in MINA 3.0, the IoBuffer class will probably be removed, or
> refactored, as we don't really need all this mechanism, but this is
> more something we should discuss on the ML.
> - the class hierarchy is probably overkilling. We should have a
> IoBuffer interface, an AbstractIoBuffer abstract class, move the
> IoBufferWrapper and ProxyHandshakeIoBuffer classes to be subclasses of
> the AbstractIoBuffer class.
> - I'm not convinced that using the CachedBuffer brings any speed
> improvement. Trying to be smarter than the VM is most certainly a lost
> of time... IMO, this class should be removed. "
>
> I understand the performance of the new IoBuffer 3.0 that will have a
> list of ByteBuffers
> (http://cwiki.apache.org/confluence/display/MINA/MINA+3.0+design) so
> will provide a continuous stream of bytes that spans several
> ByteBuffers (zero copy on auto-resize, just add a one more ByteBuffer
> to the list in the IoBuffer), but when writing the list of ByteBuffers
> out on the network, how do you want to do it? I understand that the
> gathered write has some problems in JDK, I heard some rumors about
> this, for example:
> http://lists.jboss.org/pipermail/netty-users/2009-July/000939.html
> Though I did not tested personally if there is a problem with gathered writes.
>   
IMO, there are two different cases. The zero copy things is really 
important when transfering huge block of data, as it saves CPU cycles 
(http://www.ibm.com/developerworks/library/j-zerocopy/), but if there is 
any transformation (encoding, for instance), I'm not sure that it's a 
real problem compared to the complexity it carries.

My personnal guess is that if someone has to implement a file transfert 
server, then he should have a way to use a direct copy to he clients. 
Otherwise, let's assume that a list of buffer will be more efficient 
than an extented buffer which will be created and copied for every bit 
of data coming in.


Mime
View raw message