[MPICH] An idle communication process use the same CPU as computation process on multi-core chips
Darius Buntinas
buntinas at mcs.anl.gov
Tue Sep 18 13:23:15 CDT 2007
From the discussion on lkml and the fact that they see programs that
use sched_yield() this way as "fundamentally broken", it seems that this
patch is only temporary, and eventually the pre-2.6.22 kernel behavior
won't be supported.
-d
On 09/18/2007 12:45 PM, Bob Soliday wrote:
> Well I reported the bug and it turns out they already have a patch for
> it that will be included in a future release so that it will be possible
> to emulate the old scheduler.
>
> https://bugzilla.redhat.com/show_bug.cgi?id=295071
>
> http://lkml.org/lkml/2007/9/14/157
>
> --Bob
>
> Bob Soliday wrote:
>> It turns out the problem is not related to the number of cores. Only
>> the newest versions of the Fedora 7 kernel show the problem. I think
>> it is related to the CFS scheduler in these kernels.
>>
>> When I run one slave and one master on the same core with
>> kernel-2.6.21-1.3194 using Darius's slave code I see the slave task
>> use 100% of the CPU and see the same timing values as when I run the
>> slave on a different core.
>>
>> When I do the same test with kernel-2.6.22.4-65 or kernel-2.6.22.5.76
>> the timing values double as the slave can only get 50% of the CPU time
>> when on the same core.
>>
>> --Bob
>>
>> Darius Buntinas wrote:
>>
>>>
>>> I can verify that I saw the same problem Yusong did when starting the
>>> master first on a dual quadcore machine. But assigning each slave to
>>> its own core (using taskset) fixed that.
>>>
>>> Interestingly, when there are less than 8 slaves, top shows that the
>>> master has 100% usage (when top is in "irix mode", and 12.5% (1/8)
>>> when not in irix mode). When I have 8 slaves, the usage of the
>>> master process goes to 0.
>>>
>>> Yusong, I'm betting that if you set the cpu affinity for the slaves,
>>> you'll see no impact of the master on the slaves. Can you try that?
>>>
>>> e.g.,:
>>> ./master &
>>> for i in `seq 0 3` ; do taskset -c $i ./slave & done
>>>
>>> -d
>>>
>>> On 09/17/2007 02:31 AM, Sylvain Jeaugey wrote:
>>>
>>>> This seems to be the key of the problem. When the master is launched
>>>> before others, it takes one CPU and this won't change until for any
>>>> scheduling reason he comes to share its CPU (with a slave). It then
>>>> falls to 0% and we're saved.
>>>>
>>>> So, to conduct you experiment, you definetely need to taskset your
>>>> slaves. Just launch them with
>>>> taskset -c <cpu> ./slave (1 process per cpu)
>>>> or use the -p option of taskset to do it after launch and ensure
>>>> that each slave _will_ take one CPU. Thus, the master will be
>>>> obliged to share the cpu with others and sched_yield() will be
>>>> effective.
>>>>
>>>> Sylvain
>>>>
>>>> On Sun, 16 Sep 2007, Yusong Wang wrote:
>>>>
>>>>> I did the experiments on four types of muti-core chips (2
>>>>> dual-core, 1 quad-core and 1 eight-core). All of my tests shows
>>>>> the idle master process has a big impact on the other slave
>>>>> processes except for the test of the quad-core, in which I found
>>>>> the order does matter: when the master was launched after the slave
>>>>> processes were launched, there is no affect, while if the master
>>>>> started first, two slaves processes would go to the same core and
>>>>> cause the two processes to slow down significantly than others.
>>>>>
>>>>> Yusong
>>>>>
>>>>> ----- Original Message -----
>>>>> From: Darius Buntinas <buntinas at mcs.anl.gov>
>>>>> Date: Friday, September 14, 2007 12:55 pm
>>>>> Subject: Re: [MPICH] An idle communication process use the same CPU
>>>>> as computation process on multi-core chips
>>>>>
>>>>>>
>>>>>> It's possible that different versions of the kernel/os/top compute
>>>>>> %cpu
>>>>>> differently. "CPU utilization" is really a nebulous term. What
>>>>>> you
>>>>>> really want to know is whether the master is stealing significant
>>>>>> cycles
>>>>>> from the slaves. A test of this would be to replace Sylvain's
>>>>>> slave
>>>>>> code with this:
>>>>>>
>>>>>> #include <sys/time.h>
>>>>>> int main() {
>>>>>> while (1) {
>>>>>> int i;
>>>>>> struct timeval t0,t1;
>>>>>> double usec;
>>>>>>
>>>>>> gettimeofday(&t0, 0);
>>>>>> for (i = 0; i < 100000000; ++i)
>>>>>> ;
>>>>>> gettimeofday(&t1, 0);
>>>>>>
>>>>>> usec = (t1.tv_sec * 1e6 + t1.tv_usec) - (t0.tv_sec * 1e6 +
>>>>>> t0.tv_usec);
>>>>>> printf ("%8.0f\n", usec);
>>>>>> }
>>>>>> return 0;
>>>>>> }
>>>>>>
>>>>>> This will repeatedly time the inner loop. On an N core system, run
>>>>>> N of
>>>>>> these, and look at the times reported. Then start the master and
>>>>>> see if
>>>>>> the timings change. If the master does steal significant cycles
>>>>>> from
>>>>>> the slaves, then you'll see the timings reported by the slaves
>>>>>> increase.
>>>>>> On my single processor laptop (fc6, 2.6.20), running one slave, I
>>>>>> see
>>>>>> no impact from the master.
>>>>>>
>>>>>> Please let me know what you find.
>>>>>>
>>>>>> As far as slave processes hopping around on processors, you can set
>>>>>> processor affinity ( http://www.linuxjournal.com/article/6799 has a
>>>>>> good
>>>>>> description) on the slaves.
>>>>>>
>>>>>> -d
>>>>>>
>>>>>> On 09/14/2007 12:11 PM, Bob Soliday wrote:
>>>>>>
>>>>>>> Sylvain Jeaugey wrote:
>>>>>>>
>>>>>>>> That's unfortunate.
>>>>>>>>
>>>>>>>> Still, I did two programs. A master :
>>>>>>>> ----------------------
>>>>>>>> int main() {
>>>>>>>> while (1) {
>>>>>>>> sched_yield();
>>>>>>>> }
>>>>>>>> return 0;
>>>>>>>> }
>>>>>>>> ----------------------
>>>>>>>> and a slave :
>>>>>>>> ----------------------
>>>>>>>> int main() {
>>>>>>>> while (1);
>>>>>>>> return 0;
>>>>>>>> }
>>>>>>>> ----------------------
>>>>>>>>
>>>>>>>> I launch 4 slaves and 1 master on a bi dual-core machine. Here
>>>>>>
>>>>>>
>>>>>> is the
>>>>>>
>>>>>>>> result in top :
>>>>>>>>
>>>>>>>> PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+
>>>>>>
>>>>>>
>>>>>> COMMAND>> 12361 sylvain 25 0 2376 244 188 R 100 0.0
>>>>>> 0:18.26 slave
>>>>>>
>>>>>>>> 12362 sylvain 25 0 2376 244 188 R 100 0.0 0:18.12 slave
>>>>>>>> 12360 sylvain 25 0 2376 244 188 R 100 0.0 0:18.23 slave
>>>>>>>> 12363 sylvain 25 0 2376 244 188 R 100 0.0 0:18.15 slave
>>>>>>>> 12364 sylvain 20 0 2376 248 192 R 0 0.0 0:00.00 master
>>>>>>>> 12365 sylvain 16 0 6280 1120 772 R 0 0.0 0:00.08 top
>>>>>>>>
>>>>>>>> If you are seeing 66% each, I guess that your master is not
>>>>>>>> sched_yield'ing as much as expected. Maybe you should look at
>>>>>>>> environment variables to force yield when no message is
>>>>>>
>>>>>>
>>>>>> available, and
>>>>>>
>>>>>>>> maybe your master isn't so idle after all and has message to
>>>>>>
>>>>>>
>>>>>> send
>>>>>>
>>>>>>>> continuously, thus not yield'ing.
>>>>>>>>
>>>>>>>
>>>>>>> On our FC5 nodes with 4 cores we get similar results. But on our
>>>>>>
>>>>>>
>>>>>> FC7
>>>>>>
>>>>>>> nodes with 8 cores we don't. The kernel seems to think that all 9
>>>>>>
>>>>>>
>>>>>> jobs
>>>>>>
>>>>>>> require 100% and they end up jumping from one core to another.
>>>>>>
>>>>>>
>>>>>> Often the
>>>>>>
>>>>>>> master job is left on it's own core while two slaves run on another.
>>>>>>>
>>>>>>> PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ P
>>>>>>
>>>>>>
>>>>>> COMMAND> 20127 ywang25 20 0 106m 22m 4168 R 68 0.5
>>>>>> 0:06.84 0 slave
>>>>>>
>>>>>>> 20131 ywang25 20 0 106m 22m 4184 R 73 0.5 0:07.26 1 slave
>>>>>>> 20133 ywang25 20 0 106m 22m 4196 R 75 0.5 0:07.49 2 slave
>>>>>>> 20129 ywang25 20 0 106m 22m 4176 R 84 0.5 0:08.44 3 slave
>>>>>>> 20135 ywang25 20 0 106m 22m 4176 R 73 0.5 0:07.29 4 slave
>>>>>>> 20132 ywang25 20 0 106m 22m 4188 R 70 0.5 0:07.04 4 slave
>>>>>>> 20128 ywang25 20 0 106m 22m 4180 R 78 0.5 0:07.79 5 slave
>>>>>>> 20130 ywang25 20 0 106m 22m 4180 R 74 0.5 0:07.45 6 slave
>>>>>>> 20134 ywang25 20 0 106m 24m 6708 R 80 0.6 0:07.98 7
>>>>>>
>>>>>>
>>>>>> master>
>>>>>>
>>>>>>> 20135 ywang25 20 0 106m 22m 4176 R 75 0.5 0:14.75 0 slave
>>>>>>> 20132 ywang25 20 0 106m 22m 4188 R 79 0.5 0:14.96 1 slave
>>>>>>> 20130 ywang25 20 0 106m 22m 4180 R 99 0.5 0:17.32 2 slave
>>>>>>> 20129 ywang25 20 0 106m 22m 4176 R 100 0.5 0:18.44 3 slave
>>>>>>> 20127 ywang25 20 0 106m 22m 4168 R 75 0.5 0:14.36 4 slave
>>>>>>> 20133 ywang25 20 0 106m 22m 4196 R 96 0.5 0:17.09 5 slave
>>>>>>> 20131 ywang25 20 0 106m 22m 4184 R 78 0.5 0:15.02 6 slave
>>>>>>> 20128 ywang25 20 0 106m 22m 4180 R 99 0.5 0:17.70 6 slave
>>>>>>> 20134 ywang25 20 0 106m 24m 6708 R 100 0.6 0:17.97 7
>>>>>>
>>>>>>
>>>>>> master>
>>>>>>
>>>>>>> 20130 ywang25 20 0 106m 22m 4180 R 87 0.5 0:25.99 0 slave
>>>>>>> 20132 ywang25 20 0 106m 22m 4188 R 79 0.5 0:22.83 0 slave
>>>>>>> 20127 ywang25 20 0 106m 22m 4168 R 75 0.5 0:21.89 1 slave
>>>>>>> 20133 ywang25 20 0 106m 22m 4196 R 98 0.5 0:26.94 2 slave
>>>>>>> 20129 ywang25 20 0 106m 22m 4176 R 100 0.5 0:28.45 3 slave
>>>>>>> 20135 ywang25 20 0 106m 22m 4176 R 74 0.5 0:22.12 4 slave
>>>>>>> 20134 ywang25 20 0 106m 24m 6708 R 98 0.6 0:27.73 5
>>>>>>
>>>>>>
>>>>>> master> 20128 ywang25 20 0 106m 22m 4180 R 90 0.5
>>>>>> 0:26.72 6 slave
>>>>>>
>>>>>>> 20131 ywang25 20 0 106m 22m 4184 R 99 0.5 0:24.96 7 slave
>>>>>>>
>>>>>>> 20133 ywang25 20 0 91440 5756 4852 R 87 0.1 0:44.20 0 slave
>>>>>>> 20132 ywang25 20 0 91436 5764 4860 R 80 0.1 0:39.32 0
>>>>>>
>>>>>>
>>>>>> slave
>>>>>>
>>>>>>> 20134
>>>>>>> ywang25 20 0 112m 36m 11m R 96 0.9 0:47.35 5 master
>>>>>>> 20129 ywang25 20 0 91440 5736 4832 R 91 0.1 0:46.84 1 slave
>>>>>>> 20130 ywang25 20 0 91440 5748 4844 R 83 0.1 0:43.07 3 slave
>>>>>>> 20131 ywang25 20 0 91432 5744 4840 R 84 0.1 0:41.20 4 slave
>>>>>>> 20134 ywang25 20 0 112m 36m 11m R 96 0.9 0:47.35 5
>>>>>>
>>>>>>
>>>>>> master> 20128 ywang25 20 0 91432 5752 4844 R 93 0.1
>>>>>> 0:45.36 5 slave
>>>>>>
>>>>>>> 20127 ywang25 20 0 91440 5724 4824 R 94 0.1 0:40.56 6 slave
>>>>>>> 20135 ywang25 20 0 91440 5736 4832 R 92 0.1 0:39.75 7 slave
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>
>>
>
More information about the mpich-discuss
mailing list