[IPython-dev] Kernel-client communication
Mon Aug 30 03:51:44 CDT 2010
> - I see the possibilities of distributed computing by connecting multiple
> > kernels to a single client. However, I don't get why you would want to
> > connect multiple clients to a single kernel at the same time?
> Collaboration: you're working on a problem and would like to discuss
> it with a colleague. She opens a frontend pointed to your same kernel
> and voilà, you're sharing a set of data and code you can both work on,
> type code into, make plots from, etc. Think of it like desktop
> sharing but for code.
> Ad-hoc monitoring of a computation: you have a kernel you left in the
> office running a long computation. From the bar, you log in with your
> Android frontend, view the information it's printing, and log out
> knowing that everyting is OK. Or you stop it when you realize
> something went crazy.
> Ad-hoc continuation of work: you go home for the day and leave a
> session open at work. All of a sudden you have an idea and would like
> to test it, but it depends on a bunch of long computations you've
> already run at work and varaiables that are sitting in that session.
> No problem, just connect to it, try something out and disconnect again
> when satisfied.
> Monitoring: you can set up a 'read-only' client that monitors a kernel
> and publishes its output somewhere (logs, http, sms, whatever).
> There's plenty more, I'm sure. These are just a few that quickly come to
Ah right. Although I'm not sure how often one would use such this in
practice, it's certainly a nice feature, and seems to open op a range of
possibilities. I can imagine this requirement makes things considerably
harder to implement, but since you're designing a whole new protocol from
scratch, it's probably a good choice to include it now.
> > - I saw an example in which you're kind of going towards a
> > type of UI. Is this what you're really aiming at, or is this one possible
> > front end? I'm asking because IEP has more of a Matlab kind of UI, with
> > editor from which the user can run code (selected lines or cells: code
> > between two lines starting with two ##'s). Would that be compatible with
> > kernel you're designing?
> Absolutely! We want *both* types of interface. Evan's frontend is
> more of a terminal widget that could be embedded in an IDE, while
> Gerardo's has more the feel of a Qt-based notebook. And obviously as
> soon as an HTTP layer is written, something like the Sage notebook
> becomes the next step. Several of us are long-time Mathematica users
> and use Sage regularly, so those interfaces have obviously shaped our
> views a lot. But what we're trying to build is a *protocol* and
> infrastructure to make multiple types of client possible.
> - About the heartbeat thing to detect whether kernels are still alive. I
> > a similar concept in the channels module. I actually never realized that
> > this would fail if Python is running extension code. However, I do run
> > Cython code that takes about a minute to run without problems. Is that
> > because it's Cython and the Python interpreter is still involved? I'll do
> > some test running Cython and C code next week.
> The question is whether your messaging layer can continue to function
> if you have a long-running computation that's not in Python. You can
> easily see that by just calling a large SVD, eigenvalue decomposition
> or FFT from scipy, things that are easy to make big and that are
> locked inside some Fortran routine for a long time. In that scenario,
> your program will not touch the python parts until the Fortran (or
> pure C) finish. Whether that's detrimental to your overall app or not
> depends on how the other parts handle one component being unresponsive
> for a while.
> In our case obviously the kernel itself remains unresponsive, but the
> important part is that the networking doesn't suffer. So we have
> enough information to take action even in the face of an unresponsive
I'm quite a new to networking, so sorry for if this sounds stupid: Other
than the heartbeat stuff not working, would it also have other effects? I
mean, data can not be send or received, so would maybe network buffers
overflow or anything?
Further, am I right that the heartbeat is not necessary when communicating
between processes on the same box using 'localhost' (since some network
layers are bypassed)? That would give a short term solution for IEP.
> > Since I think its interesting to see that we've taking rather different
> > approaches to do (more or less) the same thing, I'll share some
> > on what I do in IEP:
> > I use one Channels instance from the channels.py module, which means all
> > communication goes over one socket. However, I can use as many as 128
> > different channels each way. Instead of a messaging format, I use a
> > for each task. By the way, I'm not saying my method is better; yours is
> > probably more "scalable", mine requires no/little message processing. So
> > from the kernel's perspective, I have one receiving channel for stdin,
> > sending for stdout and stderr, one receiving for control (mostly
> > at the moment) and one sending for status messages (whether busy/ready,
> > debug info). Lastly there's one receiving and one sending channel for
> > introspection requests and responses.
> Interesting... But I imagine each channel requires a socket pair,
> right? In that case then you'll definitely have problems if you want
> to have hundreds/thousasnds of kernels, as you'll eventually run out
> of ports for connections. Since that's a key part of ipython, we need
> a design that scales well in that direction from the get-go. But I
> see how your approach provides you with important benefits in certain
No, that's the great thing! All channels are multiplexed over the same
socket pair. When writing a message to a channel, it is put in a queue,
adding a small header to indicate the channel id. There is a single thread
that sends and receives messages over the socket. It just pops the messages
from the queue and sends them to the other side. At the receiver side, the
messages are distributed to the queue corresponding to the right channel. So
there's one 'global' queue on the sending side and one queue per channel on
the receiver side.
> > To receive code, sys.stdin is replaced with a receivingChannel from
> > channels.py, which is non-blocking. The readline() method (which is what
> > raw_input() uses) *is* blocking, so that raw_input() behaves
> > The remote process runs an interpreter loop. Each iteration the
> > checks (non-blocking) the stdin for a command to be run. If there is, it
> > does so using almost the same code in code.py. Next (if required) process
> > GUI events. Next produce prompt if necessary, and send status. In another
> > thread, there is a loop that listens for introspection requests
> > (auto-completion, calltips, docs).
> What happens if the user wants to execute in the remote process which
> itself calls raw_input()? For example, can one call pdb in
> post-mortem mode in the remote process?
All goes well. Calling code that uses raw_input(), simply uses the
sys.stdin.readline() method. Where the sys.stdin is simply a
ReceivingChannel instance from channels.py.
In any case, thanks a lot for your interest!
> Especially now with your license change, it would be wonderful if the
> two projects could collaborate more closely.
I'm looking forward to it!
PS: I changed the topic name of this e-mail to something that better
represents what we're discussing :)
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the IPython-dev