RFR: 8247469: getSystemCpuLoad() returns -1 on linux when some offline cpus are present and cpusets.effective_cpus is not available

[Baesken, Matthias]

Hello, please review  the  following  change .
We have a Linux machine  where 
           OperatingSystemMXBean mbean = (com.sun.management.OperatingSystemMXBean) ManagementFactory.getOperatingSystemMXBean();
             double load = mbean.getSystemCpuLoad();
 returns -1  ; 

Reason is that  there are offline CPUs  (48 configured ,  32 online ). Additionally   cpusets.effective_cpus  is not available on that Linux system .


Bug/webrev :

https://bugs.openjdk.java.net/browse/JDK-8247466

http://cr.openjdk.java.net/~mbaesken/webrevs/8247469.0/


Thanks, 
        Matthias


-----Original Message-----
From: Bob Vandette <bob.vandette at oracle.com> 
Sent: Freitag, 12. Juni 2020 15:02
To: Baesken, Matthias <matthias.baesken at sap.com>
Cc: daniil.x.titov at oracle.com
Subject: Re: getCpuLoad() / getSystemCpuLoad() returns -1 on linux when some offline cpus are present and cpusets.effective_cpus is not available

I looks like there are two problems here:

1. containerMetrics.getCpuSetCpus().length returns the online CPUs but getHostConfiguredCpuCount0() returns the total number of CPUs including offline ones.

One solution might be to add a getHostOnlineCpuCount0() function.

2. If getEffectiveCpuSetCpus is not available then we should use getCpuSetCpus.

Bob.


> On Jun 12, 2020, at 6:43 AM, Baesken, Matthias <matthias.baesken at sap.com> wrote:
> 
> Hello, I noticed the following on  one of our Linux machines :
>  
>             OperatingSystemMXBean mbean = (com.sun.management.OperatingSystemMXBean) ManagementFactory.getOperatingSystemMXBean();
>             double load = mbean.getSystemCpuLoad();
>  
> returns -1  ; this seems to be related to   “8226575: OperatingSystemMXBean should be made container aware”   .
>  
> This machine has the following “special  features” 
>  
> - a few CPUs are offline (means the configured cpus are 48 but the online cpus are 32) so 
>  
>  
>     private boolean isCpuSetSameAsHostCpuSet() {
>         if (containerMetrics != null && containerMetrics.getCpuSetCpus() != null) {
>             return containerMetrics.getCpuSetCpus().length == getHostConfiguredCpuCount0();
>         }
>         return false;
>     }
>  
> Returns false
> 
> - the machine does not have cpusets.effective_cpus    (not all Linux machines have it )
> 
> In this case getSystemCpuLoad() / getCpuLoad() returns -1 (because it checks that 48 != 32, and next it checks for cpusets.effective_cpus which is not present ).
> 
> See the coding at :
> https://hg.openjdk.java.net/jdk/jdk/file/bdc14b8d31ff/src/jdk.management/unix/classes/com/sun/management/internal/OperatingSystemImpl.java#l136
>  
>                 // If the cpuset is the same as the host's one there is no need to iterate over each CPU
>                 if (isCpuSetSameAsHostCpuSet()) {
>                     return getCpuLoad0();
>                 } else {
>                     int[] cpuSet = containerMetrics.getEffectiveCpuSetCpus();
>                     if (cpuSet != null && cpuSet.length > 0) {
>                         double systemLoad = 0.0;
>                         for (int cpu : cpuSet) {
>                             double cpuLoad = getSingleCpuLoad0(cpu);
>                             if (cpuLoad < 0) {
>                                 return -1;
>                             }
>                             systemLoad += cpuLoad;
>                         }
>                         return systemLoad / cpuSet.length;
>                     }
>                     return -1;
>                 }
>  
>  
> Could  we better a)  return  the native getCpuLoad0();  in this case   or  b)  use the  available  containerMetrics.getCpuSetCpus(); when getEffectiveCpuSetCpus();
> Gives an empty array   (btw.   getCpuSetCpus()   returns  on this machine the online cpus = 0,31 = 32 ) ?
>  
> I opened
>  
> https://bugs.openjdk.java.net/browse/JDK-8247469
>  
> to track this   (I see the issue  in jdk/jdk   but   it seems  it came also  to  oracle jdk8u261,  which the July update ).
>  
> Best regards, Matthias