Re: Remote asynchronous processes

[Felipe Lema]

Hello, there (I hope I'm replying correctly)

Answering as I'm interested in this matter, although I'm not a formal
Emacs maintainer myself.

I've tried this patch for async processes and, given that I deal with
far away servers that have a significant delay in answering queries, I
find this most useful. It cuts starting processes from about 5s to less
than 1s.

While this looks promising, I was also wondering about an alternative
solution.

I understand that users have requested dealing with very specific
setups in remote servers and that currently tramp-sh deals with them:
It sends some commands to the server and adjusts the current "command
to be sent".

Instead of sending-and-waiting-for commands, can't these results be
stored in memory? That way tramp-sh can construct the "command to be
sent" /without/ doing queries. If necessary, the results stored in
memory can be updated frequently in the background. Maybe they can even
be persisted for future use (using the `persist` package).

If this is not feasible, then I can manage with this "non-checks-just-
run-this-command-as-is" approach.

Gards

Michael Albinus <address@hidden> wrote:

>Hi,
>
>> On the Tramp ML, there is a discussion about performance of remote
>> asynchronous processes. start-file-process / make-process take too
much
>> time to finish.
>>
>> One of the reasons is, that Tramp opens first a shell on the remote
>> host, performs sanity checks, and runs the command after that. Well,
I
>> cannot change this in general; the sanity checks have been added due
to
>> feedback from users.
>>
>> One idea to change the situation is, to remove all sanity checks
from
>> make-process. That is, if a user has a default directory
>> "/ssh:user@host:/path/to/dir", and if he calls
>>
>> (make-process
>>  :name "test"
>>  :buffer (current-buffer)
>>  :command '("cmd")
>>  :file-handler t))
>>
>> this is translated directly into
>>
>> ssh -l user -o ControlMaster=auto -o ControlPath='tramp.%C' \
>>   -o ControlPersist=no host "cd /path/to/dir; cmd"
>>
>> This would improve performance significantly. The drawback is, that
>> Tramp does not perform convenience checks, like password handling.
>
>I have played with this idea, and the output is the appended file
>tramp-make-process.el. It changes the make-process implementation of
>Tramp for all methods defined in tramp-sh.el (like "ssh") and
>tramp-adb.el.
>
>This is a proof-of-concept, and shouldn't be used for production. Read
>the commentary in the file for limitations.
>
>However, the speed optimization is remarkable. I've tested it with the
>following code snippet:
>
>--8<---------------cut here---------------start------------->8---
>(let ((tramp-verbose 0)
>      (default-directory "/ssh::/"))
>  ;; Fill the caches.
>  (start-file-process "" nil "true")
>  ;; Run benchmark.
>  (benchmark-run 10
>    (start-file-process "" nil "true")))
>--8<---------------cut here---------------end--------------->8---
>
>In the default case, the result is
>
>--8<---------------cut here---------------start------------->8---
>(3.623666842 80 1.183512312)
>--8<---------------cut here---------------end--------------->8---
>
>If tramp-make-process.el is loaded, the result is
>
>--8<---------------cut here---------------start------------->8---
>(0.022762177 0 0.0)
>--8<---------------cut here---------------end--------------->8---
>
>Similar results, if I use "/adb::/" as default directory:
>
>--8<---------------cut here---------------start------------->8---
>(4.599374061 2 0.03429497299999973)
>--8<---------------cut here---------------end--------------->8---
>
>vs
>
>--8<---------------cut here---------------start------------->8---
>(0.013003183 0 0.0)
>--8<---------------cut here---------------end--------------->8---
>
>Comments?
>
>Best regards, Michael.
>