[
https://issues.apache.org/jira/browse/MESOS-5589?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15323461#comment-15323461
]
Rinaldo DiGiorgio commented on MESOS-5589:
------------------------------------------
g++ Program
#include <iostream>
#include <sys/time.h>
#include <ctime>
#include <unistd.h>
using namespace std;
int main() {
for (int i = 0; i < 100; i++) {
cout << "i is " << i << endl;
struct timeval tp;
gettimeofday(&tp, NULL);
long int start = tp.tv_sec * 1000 + tp.tv_usec / 1000;
usleep(10*1000);
gettimeofday(&tp, NULL);
long int end = tp.tv_sec * 1000 + tp.tv_usec / 1000;
long diff = end - start;
cout<<"Elapsed time is : "<< diff <<" ms "<<endl;
}
return 0;
}
Sample Output
Received SUBSCRIBED event
Subscribed executor on 192.168.0.128
Received LAUNCH event
Starting task SleepLatency
Forked command at 1922
sh -c 'cd /tmp && ./sleep'
i is 0
Elapsed time is : 110 ms
i is 1
Elapsed time is : 100 ms
i is 2
Elapsed time is : 104 ms
i is 3
Running the same application without mesos is what is expected by convention --
not by spec.
Elapsed time is : 12 ms
i is 93
Elapsed time is : 12 ms
i is 94
Elapsed time is : 10 ms
i is 95
Elapsed time is : 11 ms
i is 96
Elapsed time is : 13 ms
i is 97
Elapsed time is : 10 ms
i is 98
Elapsed time is : 12 ms
i is 99
> mesos OSX Agent creates an environment where scheduling latency is 10x worse
> than without mesos
> -----------------------------------------------------------------------------------------------
>
> Key: MESOS-5589
> URL: https://issues.apache.org/jira/browse/MESOS-5589
> Project: Mesos
> Issue Type: Bug
> Affects Versions: 1.0.0, 0.28.1, 0.28.2
> Reporter: Rinaldo DiGiorgio
>
> Two applications were written, one in Java and another in C++ to measure
> latency after a sleep. In both cases the latency is often 10X the time
> without running in the mesos agent.
> This behavior was verified on mesos1.0.0, 0.28.2 and 0.28.1
> A loop is executed that delays for 10ms, the actual amoutn of time is between
> 10 and several hundred ms. There is no gurantee of this time being in a
> range. On Linux it works well, 1-3 ms more on average. On OSX it is anywhere
> from 2 to several hundred ms. Can we make the performance equivalent to
> Linux for the same type of machine. Some members of the mesos staff have
> reported that they do not see this issue. I have included all the version
> numbers
> OSX 10.11.4
> java 1.8 and 1.7
> g++
> Mesos 1.0.0 and 0.28.2
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
