Allocation of threads on a dual CPU system?

lhnewsgrps · External Since: Jul 31, 2007 Posts: 1

I'm working on an RHEL 3 box with 2 CPUs. If a process is started on
one processor and spins up 200 POSIX threads, will the threads be
distributed among the 2 CPUs or just run on the CPU that the process
started on?

Is there a way to for a thread to obtain the number of the CPU its
running on? I'm scanning through the POSIX documentation but haven't
found it yet, if it exists.

Thanks.

Les

Olivier Croquette · External Since: Aug 20, 2006 Posts: 6

lhnewsgrps RemoveThis @yahoo.com wrote, On 31/07/07 22:47:
> I'm working on an RHEL 3 box with 2 CPUs. If a process is started on
> one processor and spins up 200 POSIX threads, will the threads be
> distributed among the 2 CPUs or just run on the CPU that the process
> started on?

In the general case, the threads will run on any process and may jump
from one to another at anytime.

If you want them to stick, use sched_setaffinity().
Note that:
- the affinity is inherited by the children of a process
- the kernel does a good job at load-balancing, so using the affinity is
not necessarily a good idea.

> Is there a way to for a thread to obtain the number of the CPU its
> running on? I'm scanning through the POSIX documentation but haven't
> found it yet, if it exists.

It's not POSIX, but it's available under Linux: gethostid

http://developer.apple.com/documentation/Darwin/Reference/ManPages/man...ethosti

Note that the result of this call may change several times per second
and may have become inaccurate before it even returns.

Paul Pluzhnikov · External Since: Nov 07, 2004 Posts: 355

Olivier Croquette <ocroquette.TakeThisOut@ocroquette.free.fr> writes:

> lhnewsgrps.TakeThisOut@yahoo.com wrote, On 31/07/07 22:47:

>> Is there a way to for a thread to obtain the number of the CPU its
>> running on?

Only kernel knows that, so it would have to be a system call.
I do not believe there is such a system call.

Additional complication: the kernel keeps track of physical processor
id, and core id witin that processor. What is "the number of the CPU"
a thread is running on, on a dual-processor system where every
processor is quad-core?

> It's not POSIX, but it's available under Linux: gethostid

Gethostid returns an arbitrary number contained in /etc/hostid file,
or lacking that file, an IPv4 address of the host.

What does this have to do with the numer of the CPU this thread is
currently running on?

> Note that the result of this call may change several times per second
> and may have become inaccurate before it even returns.

I believe you mis-understood what gethostid returns.

Cheers,
--
In order to understand recursion you must first understand recursion.
Remove /-nsp/ for email.

ellis · External Since: Apr 02, 2004 Posts: 193

In article <1185914841.833028.254020.TakeThisOut@x40g2000prg.googlegroups.com>,
<lhnewsgrps.TakeThisOut@yahoo.com> wrote:

>Is there a way to for a thread to obtain the number of the CPU its
>running on?

Why would you care?

--
http://www.spinics.net/lists/kvm-devel/

Olivier Croquette · External Since: Aug 20, 2006 Posts: 6

Paul Pluzhnikov wrote, On 1/08/07 7:19:
> Only kernel knows that, so it would have to be a system call.

There are instructions giving information about the CPU on common
platforms, so I can imagine it's more or less doable from userspace on
these ones.

> I do not believe there is such a system call.
[...]
> Gethostid returns an arbitrary number contained in /etc/hostid file,
> or lacking that file, an IPv4 address of the host.

You're right. I don't know what went through my mind at the moment I
posted that...

>> Note that the result of this call may change several times per second
>> and may have become inaccurate before it even returns.
>
> I believe you mis-understood what gethostid returns.

Well, if there was such a call, the above statement about it would be
correct Smile

Some resources on this topic:

http://sourceware.org/ml/libc-alpha/2006-06/msg00024.html

smp_processor_id() (kernel):
http://kernel.org/doc/htmldocs/kernel-hacking.html#routines-processorids

So it seems there is no /easy/ way.

David Schwartz · External Since: Jun 01, 2007 Posts: 87

On Jul 31, 1:47 pm, lhnewsg....DeleteThis@yahoo.com wrote:

> I'm working on an RHEL 3 box with 2 CPUs. If a process is started on
> one processor and spins up 200 POSIX threads, will the threads be
> distributed among the 2 CPUs or just run on the CPU that the process
> started on?

You can't say that they will, because there are too many variables.
They well might be. For example, if the other CPU was being used by a
task using a privileged scheduler, these ordinary threads would not be
scheduled on it unless that other task voluntarily yielded.

> Is there a way to for a thread to obtain the number of the CPU its
> running on? I'm scanning through the POSIX documentation but haven't
> found it yet, if it exists.

Nope, because the kernel is free to deschedule or migrate the thread
at any time. So such a function would be impossible to write except in
conjunction with some way to lock a thread to a particular processor.

You could, for example, tell the scheduler to only schedule you on a
particular CPU and then yield. On some OSes (I think Linux too, but
I'd have to check), you are guaranteed only to resume on a CPU that
you are currently allowed to run on.

DS

David Schwartz · External Since: Jun 01, 2007 Posts: 87

Interestingly, recent versions of Linux (with glibc-2.6) have
'sched_getcpu'. It's not particularly useful, but it does exist.

Here's some sample code:

#include <sched.h>
#include <stdio.h>
#include <pthread.h>

void *tf(void *t)
{
int i, j, q, cpu;
for(i=0; i<25600; i++)
for(j=0; j<25600; j++)
{ /* burn some CPU */
q>>=2;
q+=i;
q-=j;
}
cpu=sched_getcpu();
for(i=0; i<25600; i++)
for(j=0; j<25600; j++)
{ /* keep burning */
q<<=2;
q+=i;
}
printf("%d Running on CPU %d\n",
(int) t, cpu);
*(volatile int *) &q=q+1; /* don't let compiler optimize q away */
return (void *) cpu;
}

int main(void)
{ /* create 8 CPU burners and see what CPUs they run on */
pthread_t t[8];
int i;
void *r;
for(i=0; i<8; i++)
pthread_create(&t[i], NULL, tf, (void *) (i+1));
for(i=0; i<8; i++)
pthread_join(t[i], &r);
}

On my quad-core, this is typical:

8 Running on CPU 1
1 Running on CPU 1
4 Running on CPU 3
3 Running on CPU 3
5 Running on CPU 2
6 Running on CPU 0
2 Running on CPU 2
7 Running on CPU 0

DS