On Sun, Apr 20, 2014 at 1:19 AM, Guillaume Nodet <[email protected]> wrote: > 2014-04-19 20:43 GMT+02:00 Alon Bar-Lev <[email protected]>: > >> On Sat, Apr 19, 2014 at 9:15 PM, Guillaume Nodet <[email protected]> >> wrote: >> > Sshd internally uses nio2 by default, which is not based on selectors, >> but >> > non blocking operations. >> > >> > On the client part of SSHD, things are mostly asynchronous already: >> > #1 SshClient#connect returns a future on which you can set a callback >> > and that you can use to retrieve the ClientSession asynchronously >> > #2 You need to use ClientSession#addXxxIdentity and then >> > ClientSession#auth which is also asynchronous >> > #3 You then create a channel, and actually operning the channel is >> also >> > asynchronous >> > #4 Closing channels is also asynchronous >> > >> > I think the only missing part is really the streams on the ClientChannel >> > which are using InputStream and OutputStream. >> > If we replace them with an AsynchronousByteChannel, I think we would be >> > fully async. >> >> Thank you for your response, Our definition of async is very different... >> :) >> >> I do not think this module is sufficient to what I target. I see the >> number of threads created within the library core and the logic that >> is out of reach. > > >> This ssh library is great, splitting it into two logic only and >> communication layers will enable to go fully async. The logic layer >> should not have any thread. A default implementation of communication >> layer can be provided, but is optional. The difference from the world >> I coming for is that Future handling is much more complex than having >> control queue. >> > > Not sure exactly what you're talking about here. > > Afaik, the only place where the ssh layer actually create a thread in when > creating > a client ChannelSession giving an InputStream which has to be read. This > thread creation can be easily avoided by using ClientChannel#getInvertedIn() > and writing to it. > > All other threads are communication threads only and are fully controlled > by > the IoService layer which is pluggable. Both mina and nio2 implementations > use a fixed number of threads. But you can rewrite it if you need. > > I'm all for improving sshd, but I fear i'm not really seeing your points > clearly.
Thank you for the discussion, I truly appreciate that. Having a method for async input/output of data stream will be a good start within current implementation. Other than that it is a programming pattern discussion. I got the information I needed, thank you! >> >> Was just an idea, thank you for addressing. >> >> > >> > 2014-04-19 15:57 GMT+02:00 Alon Bar-Lev <[email protected]>: >> > >> >> On Sat, Apr 19, 2014 at 3:52 PM, Jon V. <[email protected]> wrote: >> >> > >> >> > NIO controls and deals with the selectors. Async IO is a part of that >> but >> >> > is not the same thing. Async io means that if a write cannot be fully >> >> > flushed. It will not block until it can be. NIO provides us the >> events to >> >> > tell us that data is available in the socket. >> >> >> >> Async IO is the ability for a single thread to perform (multiplex) IO >> >> (connect, read, write, close etc..) for multiple file descriptors. >> >> >> >> As far as I know, without NIO you cannot achieve that in Java. >> >> >> >> There is no sense in read or write without blocking if you cannot wait >> >> (vs actively poll) for an event. >> >> >> >> > On Apr 19, 2014 4:56 AM, "Alon Bar-Lev" <[email protected]> >> wrote: >> >> > >> >> > > On Sat, Apr 19, 2014 at 10:58 AM, Emmanuel Lécharny < >> >> [email protected]> >> >> > > wrote: >> >> > > > Le 4/19/14 9:45 AM, Alon Bar-Lev a écrit : >> >> > > >> On Sat, Apr 19, 2014 at 10:38 AM, Emmanuel Lécharny < >> >> > > [email protected]> wrote: >> >> > > >>> Le 4/19/14 9:13 AM, Alon Bar-Lev a écrit : >> >> > > >>>> Hi, >> >> > > >>>> >> >> > > >>>> The mission of async is to avoid having threads at all, or at >> >> least >> >> > > O(1). >> >> > > >>>> >> >> > > >>>> As you have underline internal/private low level channels for >> >> socket >> >> > > >>>> processing, and public high level channels to communicate with >> >> > > >>>> application, there should be a mechanism for library to request >> >> wake >> >> > > >>>> up for these low level channels. >> >> > > >>>> >> >> > > >>>> Another option is to avoid using sockets at all within the >> >> > > >>>> implementation and require application to manage the sockets >> and >> >> pipe >> >> > > >>>> socket data into the library. >> >> > > >>>> >> >> > > >>>> I understand this is conceptional change than what we have now, >> >> but >> >> > > >>>> this what will enable scale without abusing system threads or >> have >> >> > > >>>> nondeterministic behaviour in high load. >> >> > > >>> There are a few important things you have to know about async >> and >> >> > > threads : >> >> > > >>> - the extra cost for dealing with async connection is around >> 30%. >> >> That >> >> > > >>> all but free >> >> > > >>> - a standard system can easily deal with a few thousands of >> threads >> >> > > >>> >> >> > > >>> Now, unless you define what is "high load", I don't really see >> what >> >> > > kind >> >> > > >>> of advantage we can get with an async implementation. >> >> > > >>> >> >> > > >>> FTR, when MINA was initially created, it was because there was a >> >> need >> >> > > >>> for a system supporting potentially ten of thousands of >> >> connections. Is >> >> > > >>> that what you are targetting ? >> >> > > >> Yes, using work threads that are derived per # of CPUs, no more. >> >> > > >> I am far from the pure "Java" world... but if async IO is 30% >> >> > > >> insufficient, maybe it worth to use libssh (C) and communicate >> with >> >> it >> >> > > >> using single socket from java, delegating IO outside of java. >> >> > > > IO are already delegated outside on Java. Eveything IO related is >> >> > > > written in C and wrapped into Java class. >> >> > > > >> >> > > > The extra cost when using NIO is due to the management of >> SelectorKey >> >> > > > lists (with the various steps involved when dealing with those >> >> lists). >> >> > > > >> >> > > > All in all, when it comes to process IO, Java does not really add >> >> some >> >> > > > extra cost over a plain C implementation. It's not the same story >> >> when >> >> > > > using NIO, especially when dealing with many concurrent >> connections. >> >> > > > >> >> > > >> >> > > So I am confused... Java does not add cost to async IO, but NIO >> does? >> >> > > While NIO is the only interface to Java async IO? >> >> > > >> >> >>
