The same is still true... a Channel is pinned to an EventLoop and so it’s IO is
processed by one thread
> Am 16.02.2018 um 00:04 schrieb Sean Bright :
>
> Eric,
>
> I don't know what I can add beyond what has already been written in this
> thread. I haven't following Netty development so I don't know if this is
> still the case or not. At the time, a channel was bound to a thread in an
> EventExecutorGroup, so you could run into a situation where a group of
> channels that were all bound to the same thread would be under serviced if
> one of those tasks took a lot of CPU time. If instead you delegate to a
> standard ExecutorService, each task for a given channel might run on a
> separate thread so a single task couldn't starve all of the channels.
>
> Kind regards,
> Sean
>
>> On Friday, February 9, 2018 at 8:40:00 PM UTC-5, Eric Fulton wrote:
>> Hey sorry I know this is a very old thread, but can you explain why in you
>> example you think it would be better to use your own ExecutorService over
>> someEventExecutorGroup? What would be the difference?
>>
>>
>>
>>
>>> On Wednesday, October 9, 2013 at 7:27:06 AM UTC-7, Sean Bright wrote:
>>> Thanks Norman.
>>>
>>> So based on this, it seems that if I have a "typical" pipeline:
>>>
>>> ctx.pipeline().addLast("decoder", new MyProtocolDecoder());
>>> ctx.pipeline().addLast("encoder", new MyProtocolEncoder());
>>> ctx.pipeline().addLast(someEventExecutorGroup, "handler", new
>>> MyProtocolHandler());
>>>
>>> Using an EventExecutorGroup doesn't actually buy me anything. I/O won't be
>>> blocked, but handler execution will.
>>>
>>> I understand the point of doing all of this - if you know your methods will
>>> always be called on the same thread it reduces the synchronization
>>> complexities in the handlers - but in this model when you are dealing with
>>> hundreds of connections, any handler method that causes a delay in
>>> processing will block (num_connections / num_event_executor_threads) - 1
>>> other handlers.
>>>
>>> So it would appear that in order to get the behavior that I want (any one
>>> connection should not affect another), I would eliminate the
>>> EventExecutorGroup and would need to submit tasks to an ExecutorService
>>> that I manage myself, correct?
>>>
>>> I guess I just don't see how an EventExecutorGroup is beneficial.
>>>
>>> In any case, I love Netty. Keep up the good work!
>>>
>>>
On Tuesday, October 8, 2013 11:40:30 AM UTC-4, Sean Bright wrote:
Greetings,
I just want to validate my understanding of Netty 4's thread model
compared to Netty 3's (specifically as it applies to NIO).
With Netty 3, you would specify a boss thread pool and a worker thread
pool (all/most of the examples use Executors.newCachedThreadPool() with
the default args). ChannelHandler events would be called on the I/O
threads from the worker group. Any delay in the handler methods would
cause a delay in processing I/O for other connections. In order to
compensate for this, you could add an ExecutionHandler to the pipeline
which would cause the handler methods to be fired on an executor thread
and therefore wouldn't affect I/O.
In Netty 4, you specify a boss group and worker group, and as channels
connect they are registered to specific threads. So channel 1 will always
have it's handler events fired on thread A, channel 2 will always have
it's handler events fired on thread B. Again, any delayed processing that
occurs in the handler method will hurt I/O for other channels registered
to that worker thread. To compensate, you specify an EventExecutorGroup
so that I/O is not affected with long running tasks.
Assuming everything above is correct...
Assume that I create a DefaultEventExecutorGroup, passing 4 as the number
of threads, and assign that to my handler in the pipeline. Now, 8
channels connect:
Channel A: EventExecutor 1
Channel B: EventExecutor 2
Channel C: EventExecutor 3
Channel D: EventExecutor 4
Channel E: EventExecutor 1
Channel F: EventExecutor 2
Channel G: EventExecutor 3
Channel H: EventExecutor 4
Each channel is registered to EventExecutor thread. If Channel A in the
above example performs a long running task (say, in channelRead0), then
won't Channel E be blocked during this time? Is that correct or am I not
understanding something? If I am correct, why would I ever want to use an
EventExecutor? I feel like I would be better off using a shared Executor
directly from my handler methods (and handling thread synchronization
myself). At least in that case I wouldn't be blocking other clients.
Thank you,
Sean
>
> --
> You received this message because you are subscribed to the Google Groups
> "Netty