Maybe it would be feasible to run your ideal user space in a virtual

environment (i.e. qemu), where the guest user space thinks it can use

the CPU for itself when in fact it's just 50-75% of the actual

machine.
On Jan 14, 2008 10:27 PM, Andreas Kahari  wrote:

quoted text
>

> On 14/01/2008, Alexander Schrijver  wrote:

> > On Jan 14, 2008 2:34 PM, Andreas Kahari  wrote:

> > >

> > > On 14/01/2008, Alexander Schrijver  wrote:

> > > > On Jan 14, 2008 1:30 PM, Andreas Kahari  wrote:

> > > > > On 14/01/2008, Alexander Schrijver  wrote:

> > > > > > On Jan 14, 2008 11:52 AM, Andreas Kahari  wrote:

> > > > > > > Hi,

> > > > > > >

> > > > > > > Is there a way of limiting the amount of CPU given to a particular

> > > > > > > process or process group? For example, I would want the build of the

> > > > > > > qt4 port to use a maximum of 25% of the available CPU, leaving the CPU

> > > > > > > 75% idle if nothing else is happening on the machine.

> > > > > > >

> > > > > > > I know about 'nice', but it doesn't fulfil the criteria that the

> > > > > > > machine is left otherwise idle if nothing else runs on it.

> > > > > > >

> > > > > > > I don't have a real reason for why I would want to do this, I'm mainly

> > > > > > > curious as to if it's possible.

> > > > > [cut]

> > > > > >

> > > > > > I have never done this myself, but I believe this is possible by

> > > > > > creating a login class in /etc/login.conf and set the cputime option.

> > > > > > See login.conf(5) for a better description.

> > > > >

> > > > > Hi Alexander,

> > > > >

> > > > > I believe that the cputime resource limit will limit the maximum

> > > > > amount of CPU time that the user may use in a session, which is not

> > > > > really what I asked for. I'd like the process or process group to run

> > > > > for as long as it needs to run, but that it only ever uses a fraction

> > > > > of the CPU power.

> > > > >

> > > > > It's like limiting the network bandwidth for a particular type of

> > > > > traffic, only this is about time on the CPU.

> [cut sigs]

> > > >

> > > > Yes, you are right. It is also possible to set a priority for a

> > > > process in a login class.

> > > >

> > > > From login.conf(5)

> > > >      priority          number                  Initial priority (nice) level.

> > > >

> > > > This is not exactly what you want, but it is pretty close. I am

> > > > curious why do you want to set an exact limit and not let the

> > > > scheduler do this for you?

> > > >

> > >

> > > As I said, I don't have a good reason for wanting to do this. It just

> > > seemed like something someone might want to do. But let me dream up

> > > three examples: Sometimes firefox (or whatever program) goes a bit

> > > haywire and brings the machine to a crawl. It would be nice to limit

> > > firefox's CPU to a maximum of, say, 50% so that I'm guaranteed to have

> > > 50% of the machine to work with.

> > >

> > > Another example: Let's say I'm rebuilding the kernel, base system, and

> > > all my packages after a major update from CVS after a long time away.

> > > I'm not worried about how long this takes so I'm quite happy to run

> > > the build at 5% of the CPU while I get on with my work.

> > >

> >

> > What you describe here is exactly what you can accomplish with either

> > nice or the priority option in login.conf. Also, I am not exactly sure

> > what you mean with percent of CPU. Do you mean the difference of cpu

> > time scheduled between a 'normal' process?

>

> It is not quite the same because a process, even running at niceness

> level 20, will grab as much CPU as it can (unless it has to wait for

> data).

>

> What I mean is what I wrote in my first email: "For example, I would

> want the build of the qt4 port to use a maximum of 25% of the

> available CPU, leaving the CPU 75% idle if nothing else is happening

> on the machine."

>

> > > Third example, similar to the last one: I'm running a distributed.net

> > > or SETI-at-home client in the background, but I don't ever want it to

> > > run at 100% of the CPU, maybe because that would make the machine too

> > > noisy during the night (due to the fans).

> >

> > I think this is a different issue. I don't know how this can be

> > solved. An idea might be to underclock your cpu at night.

>

> Yes, I could "sysctl -w hw.setperf=0" or something like that on my

> SpeedStep'able CPU, but with hw.setperf=0, a process still have the

> possibility to get 100% of the CPU. It will be a slower CPU, but it's

> still not limiting the process to use only a fraction of the CPU at

> any instance in time.

>

> Again, maybe no-one has these kind of requirements in real life?

>

>

> Regards,

> Andreas

>

> --

> Andreas Kahari

> Somewhere in the general Cambridge area, UK

>

>

--

Tito Mari Francis H. EscaC1o

Computer Engineer and Free Software Proponent