Could some of this performance hit be due to the 1 to 1 heavyweight threading
model used by Linux ? Both IBM and all the Linux developers I have actually
met say this is no longer up to date with POSIX standards and a pain in the
butt, especially since it allows disconnected child processes to ignore
sig.'s after the parent has crashed. Just a thought !
Owen
On Sunday 15 April 2001 1:30 pm, you wrote:
> Bruce F. Press wrote:
> > Yes, yes, we've heard this before. It is not a satisfactory answer,
> > clearly the "idle" loop in kapm-idled could use a nice sleep(15) or
> > something!!
>
> What would be a satisfactory answer?
>
> Are you concerned because top shows the system being far busier than it
> really is?
>
> (Then maybe we need a modified top that does not count kapm-idled as
> processor usage in the "CPU States" percentage.)
>
> Do you disbelieve what you are being told (that, IIUC, the time used by
> kapm-idled is really system idle time that will transparently and
> instantaneously be applied to a real task if one exists and is ready to
> run)? If so, do you have any evidence of this -- is your system running
> slowly or more slowly than you would expect / are used to?
>
> I'm serious about these questions, not trying to be a smart ass. The
> story being told is believable to me. But, I keep hearing that Linux's
> approach of using all available RAM (not paraphrased accurately) is the
> best approach, yet in a Windows 95 system with 64 MB of memory I can
> keep 30 IE 5 windows open and get snappy response switching between
> windows, yet in a Linux (Mandrake 7.2) system with 128 MB of RAM (and
> comparable processors, identical motherboards and video cards),
> everything works much slower even with only one window open, and by the
> time I get to 15 to 20 open (Konqueror 2.0 or 2.1) windows, the system
> is like molasses.
>
> I think part of the problem is theat kde/konqueror need to learn some of
> the tricks that Windows uses. I can't describe those tricks accurately,
> but I see the results. One example: in Windows, if I create a new
> instance of IE, it appears almost instantly, and the disk doesn't make a
> peep. In Linux, if I do the same thing in konqueror, the disk starts
> chugging, and 15 to 45 seconds later the new instance of konqueror
> appears (and on the wrong desktop if I've switched desktops in the
> interim).
>
> Don't get me wrong, I want Linux to succeed, but I think some new tricks
> are needed. (Also, in Linux, if I make some wrong keystrokes, it seems
> that they are all queued up and executed (slowly) one after the other.
> In Windows, it seems that if I type (or click) the wrong command, but
> then type (or click) the right command, the initial incorrect command is
> interrupted and never completed (at least under some circumstances). I
> know I am not describing this stuff accurately or completely, but it
> sure makes a Windows system much more responsive than a Linux system.
>
> And yes, "until it crashes" -- but I have learned to watch my resource
> usage in Windows and reboot once or twice a week whether I need to or
> not, and thus rarely if ever get a crash. Yes, I would prefer not to
> have to reboot periodically, but I get more done quicker in Windows
> between reboots that I do in Linux waiting for the molasses.
>
> If you (anyone) can collaborate these stories, and help me get them to
> the right developers, it would be to the benefit of all of us. I don't
> know whether these things need to be addressed at the OS level or the
> desktop level, or someplace else, but someone needs to consider them.
>
> (And, if the desktop developers tell me they can do nothing, it all
> depends on the OS developers, I will not believe them. I might believe
> that the cleanest fix must be done at the OS level (if that's what they
> tell me), but I believe that fixes can also be done at lower levels.
> Perhaps performance can be improved by always keeping a buffer of free
> RAM large enough to immediately clone a konqueror window. At the
> desktop level, one or more such buffers can be created, even if you do
> something dumb like precreating an unused instance of konqueror. Then,
> when an instance of konqueror is requested, display this one immediately
> (with no disk chugging). Then start the disk chugging to create another
> free buffer for the next request.
>
> I know these kind of things can be done. I am not enough (or any) of a
> programmer to do them myself. I imagine all the developers are busy
> doing important things. Are they aware of and planning to implement
> techniques like these, or better?
>
> If this email has any value, please feel free to copy or quote portions
> of it to anyone, anywhere, anytime.
>
> Thanks,
> Randy Kramer
>
> > Chmouel Boudjnah wrote:
> > > SI Reasoning <[EMAIL PROTECTED]> writes:
> > > > There are good stretches of the day where my CPU spins
> > > > at around 50% or more and the process spinning is
> > > > kapm-idled. This is not a problem in 7.2.
> > >
> > > --=-=-=
> > > http://www.tux.org/lkml/#s14-1:
> > >
> > > 1.Why is kapmd using so much CPU time?
> > > (REG) Don't worry, it's not stealing valuable CPU time from
> > > other processes. It's just consuming idle cycles (normally
> > > charged to the idle task, which is displayed differently in
> > > top). Normally, when your system is idle, the system idle
> > > task is run, and this is shown as idle time (i.e. the "unused"
> > > CPU time is not charged to a specific process). With APM
> > > (Advanced Power Management), a special idle task (kapmd) is
> > > required so that greater power saving techniques can be
> > > enabled. So now, the "unused" CPU time is charged to the kapmd
> > > task instead.
> > >
> > > --=-=-=
> > >
> > > --=-=-=
> > > http://olstrans.sourceforge.net/release/OLS2000-apm/OLS2000-apm.html:
> > >
> > > In 2.2 and before, we basically had a hook into the idle loop, so that
> > > if we had APM enabled, we would just tell the BIOS that we're
> > > idle. In 2.3, Linus thought it would be a good idea if we had a
> > > separate power management idle loop, so (he) we invented the
> > > kernel APM daemon and I started getting bug reports about this
> > > process that was using all our time, called kapmd. And if you sat
> > > there just running top on a 2.3 kernel, the top process, if you're
> > > not doing anything else, will be kapmd and it will be using like
> > > 85% or 90% or 95% of your CPU time. These people were worried
> > > because it was idle: why is it using all of the time? Well
> > > actually, it's just that the time is getting accounted to that
> > > process. It's not doing anything, it's the idle loop. [26m, 12s]
> > > --=-=-=
> > >
> > > --
> > > MandrakeSoft Inc http://www.chmouel.org
> > > --Chmouel
> >
> > ---------------------------------------------------------------
> >
> > Name: brucefp.vcf
> > Part 1.2 Type: text/x-vcard
> > Encoding: 7bit
> > Description: Card for Bruce F. Press
