It would use more memory and since there is already 2x IoProcessors for cpu
cores there would be no speedup. In fact it would lead to more memory
fragmentation/cache misses and require one thread for every session.

Basically, it would be a bad idea.
On Oct 7, 2015 3:36 PM, "Марат Гайнуллин" <[email protected]> wrote:

> Dear sirs!
>     It is completely clear for me, that decoding of a buffer must be linear
> and tcp messages fragments must be decoded in a serial manner. My question
> is about different thing. Why not to decode SEPARATE sessions' data streams
> in parallel?
> 7 окт. 2015 г. 10:18 PM пользователь "Jon V." <[email protected]>
> написал:
>
> > You also cannot process a linear buffer simultaneously.  TCP is linear
> and
> > reading the messages must be linear.  After you have a list of messages
> > then you may execute them in parallel but the line decoding of the buffer
> > must be single threaded.  The decoder operates on an inbound buffer and
> > multi-threading this process will cause the buffers to complete out of
> > order.  There is also no guarantee that the inbound buffer contains a
> > complete message.  In this scenario the tail end of a message could be
> > processed before the previous part which makes the whole thing totally
> > broken without locks and locks would just eat more cpu with zero gain.
> > On Oct 7, 2015 2:53 PM, "Emmanuel Lécharny" <[email protected]> wrote:
> >
> > > Le 07/10/15 20:44, Марат Гайнуллин a écrit :
> > > > Dear sirs!
> > > >      I am happy to know, that DIRMINA-1013 has been fixed.
> > > > I have a question. I think, that while decoding TCP messages in
> > separate
> > > > sessions, server SHOULD do it in parallel. Why your last comment on
> > > > DIRMINA-934 was about not to put executor filter before decoder in
> TCP?
> > > Because doing so will put your server atrisk. Note that you already
> have
> > > more than one thread processing incoming messages (IoProcessors) and
> TCP
> > > does not guarantee you that a message will be delivered in one piece.
> > > Now, let's assume you have an exectir *before* the decoder : then for
> > > one specific session, more than one thread might be used to decode a
> > > message, which means that potentially, for fragmented messages, more
> > > than one thread will be in charge of decoding fration of a message.
> this
> > > simply does not fly...
> > >
> > > For separate sessions, there is absolutely no problem, because each
> > > session might be handled by a different IoProcessor.
> > >
> > > Performance wise, you might object that while a message is being
> > > decoded, then another session could be blocked because the thread is
> > > busy decoding (as the two sessions are handled by the same IoProcessor
> > > thread). that would be true, except that when an IoProcessor thread is
> > > busy decoding, it uses 100% of the CPU of the core it run on, so there
> > > is no 'room' for some other exacution. So having an executor before the
> > > decoder is simply useless, unless you haven't declared as many
> > > IoProcessors as you have core on your server.
> > >
> > > The only case where it's annoying is when your decoder is accessing
> > > external resources and has to wait for them : unlikely...
> > > Hope it helps !
> > >
> > >
> >
>

Reply via email to