[. On Wednesday, May 29, 2002 at 11:59:43 (-0700), Bryan Stansell wrote: ]
> Subject: Re: conserver eventually goes catatonic after SIGPIPE (on NetBSD)
>
> well, that's true, it does only initialize things one at a time, but in
> the case of '|' syntax consoles, all that requires is a fork of a
> process. could a similar mode of thinking could be used for the
> chat-based consoles?
So far I've only looked at the "chat" feature as something done
immediately after the "port" (TTY, TELNET or other TCP, pipe, etc.)
connection has been opened, and which is in some way necessary for the
sucessful ongoing functioning of the console connection. It's only done
once at initialisation time, and presumably the connection should be
closed and reopened at some later time if the chat fails.
Anything beyond that, such as periodic commands, response to certain
output at any time, etc., should be handled externally, which is what
I'm doing now to collect data from the serial ports of my UPS systems:
(printf "format\r"; sleep 2) | console -f best-ups-1 | process
> > BTW, nobody should be afraid of having one process per console port
> > either, even if they have a thousand console ports to manage. This is
> > unix, after all! (even if some implementations are a bit stupid about
> > fork(), it's not as if this application would be constantly re-forking
> > processes continuously)
>
> well, actually, maybe they should.
well, no, they shouldn't have to be afraid of one process per
connection.....
> in the past, i've seen multi-gig
> systems run low on vm because of conserver's footprint.
(hmmm.... I wonder if this could have been part of the problem I was
seeing too.....)
Normally my conserver processes run at only 100 killobytes or so or so
when they first start, but they seem to grow.... The other day I
spotted one that was 21 megabytes. Here's proof in that the child is
now over five times as big as its parent and it's done nothing but
handle bunches of "console" client connections (and this is with sixteen
ports, one for a remote conserver and the rest doing telnet connections):
$ ps -auxc | sed -n -e '1p' -e '/conserver/p'
USER PID %CPU %MEM VSZ RSS TT STAT STARTED TIME COMMAND
root 27231 0.0 0.2 492 104 ?? Is 4:42PM 0:22.80 conserver
root 27232 0.0 0.2 2548 104 ?? I 4:42PM 2:14.83 conserver
Running a few more manual client connections repeatedly, and then
disconnecting again, leaves it nearly 100KB larger already!
$ ps -auxc | sed -n -e '1p' -e '/conserver/p'
USER PID %CPU %MEM VSZ RSS TT STAT STARTED TIME COMMAND
root 27231 0.0 0.2 492 104 ?? Ss 4:42PM 0:23.82 conserver
root 27232 0.8 0.2 2640 104 ?? R 4:42PM 2:19.62 conserver
Once this leak is fixed for good there's no issue with running a
thousand or more such processes on a properly provisioned server --
remember that all modern Unix systems will share the code segment (and
most page it from the original executable too, not from swap).
(Note, I haven't started to look for this leak yet.....)
> i personally
> like the idea of being able to use a relatively low-end machine as a
> console server (i can't count the number of times i've heard of
> infrastructure machines like this left out of budgets or seen as
> unimportant by management or just added to an already loaded admin
> box).
Sure, but in this case the solution I pose will actually reduce total
resource requirements, not increase them, at least not until you get
into thousands of connections at which point you've gone way beyond the
capabilities of your low-end machine (eg. old SS2 w/32MB) anyway.
It's probably best to split the port process out into a separate binary
though -- that way the overhead of the master isn't replicated for every
connection.
Also, think about all the advantages w.r.t. configuration reloads and
long-running connections. No more need to kill clients when all you're
doing is adding or chaning only one console connection.....
--
Greg A. Woods
+1 416 218-0098; <gwoods@acm.org>; <g.a.woods@ieee.org>; <woods@robohack.ca>
Planix, Inc. <woods@planix.com>; VE3TCP; Secrets of the Weird <woods@weird.com>
> Subject: Re: conserver eventually goes catatonic after SIGPIPE (on NetBSD)
>
> well, that's true, it does only initialize things one at a time, but in
> the case of '|' syntax consoles, all that requires is a fork of a
> process. could a similar mode of thinking could be used for the
> chat-based consoles?
So far I've only looked at the "chat" feature as something done
immediately after the "port" (TTY, TELNET or other TCP, pipe, etc.)
connection has been opened, and which is in some way necessary for the
sucessful ongoing functioning of the console connection. It's only done
once at initialisation time, and presumably the connection should be
closed and reopened at some later time if the chat fails.
Anything beyond that, such as periodic commands, response to certain
output at any time, etc., should be handled externally, which is what
I'm doing now to collect data from the serial ports of my UPS systems:
(printf "format\r"; sleep 2) | console -f best-ups-1 | process
> > BTW, nobody should be afraid of having one process per console port
> > either, even if they have a thousand console ports to manage. This is
> > unix, after all! (even if some implementations are a bit stupid about
> > fork(), it's not as if this application would be constantly re-forking
> > processes continuously)
>
> well, actually, maybe they should.
well, no, they shouldn't have to be afraid of one process per
connection.....
> in the past, i've seen multi-gig
> systems run low on vm because of conserver's footprint.
(hmmm.... I wonder if this could have been part of the problem I was
seeing too.....)
Normally my conserver processes run at only 100 killobytes or so or so
when they first start, but they seem to grow.... The other day I
spotted one that was 21 megabytes. Here's proof in that the child is
now over five times as big as its parent and it's done nothing but
handle bunches of "console" client connections (and this is with sixteen
ports, one for a remote conserver and the rest doing telnet connections):
$ ps -auxc | sed -n -e '1p' -e '/conserver/p'
USER PID %CPU %MEM VSZ RSS TT STAT STARTED TIME COMMAND
root 27231 0.0 0.2 492 104 ?? Is 4:42PM 0:22.80 conserver
root 27232 0.0 0.2 2548 104 ?? I 4:42PM 2:14.83 conserver
Running a few more manual client connections repeatedly, and then
disconnecting again, leaves it nearly 100KB larger already!
$ ps -auxc | sed -n -e '1p' -e '/conserver/p'
USER PID %CPU %MEM VSZ RSS TT STAT STARTED TIME COMMAND
root 27231 0.0 0.2 492 104 ?? Ss 4:42PM 0:23.82 conserver
root 27232 0.8 0.2 2640 104 ?? R 4:42PM 2:19.62 conserver
Once this leak is fixed for good there's no issue with running a
thousand or more such processes on a properly provisioned server --
remember that all modern Unix systems will share the code segment (and
most page it from the original executable too, not from swap).
(Note, I haven't started to look for this leak yet.....)
> i personally
> like the idea of being able to use a relatively low-end machine as a
> console server (i can't count the number of times i've heard of
> infrastructure machines like this left out of budgets or seen as
> unimportant by management or just added to an already loaded admin
> box).
Sure, but in this case the solution I pose will actually reduce total
resource requirements, not increase them, at least not until you get
into thousands of connections at which point you've gone way beyond the
capabilities of your low-end machine (eg. old SS2 w/32MB) anyway.
It's probably best to split the port process out into a separate binary
though -- that way the overhead of the master isn't replicated for every
connection.
Also, think about all the advantages w.r.t. configuration reloads and
long-running connections. No more need to kill clients when all you're
doing is adding or chaning only one console connection.....
--
Greg A. Woods
+1 416 218-0098; <gwoods@acm.org>; <g.a.woods@ieee.org>; <woods@robohack.ca>
Planix, Inc. <woods@planix.com>; VE3TCP; Secrets of the Weird <woods@weird.com>