Difference between revisions of "Resource shortage"

From OpenVZ Virtuozzo Containers Wiki
Jump to: navigation, search
(Additional explanation of cpulimit on multi-core servers.)
m (cpuunits: intentation)
 
(5 intermediate revisions by 5 users not shown)
Line 90: Line 90:
 
vzctl set 123 --diskspace $(( 1048576*2 )):$(( 1153434*2 )) --save
 
vzctl set 123 --diskspace $(( 1048576*2 )):$(( 1153434*2 )) --save
 
</pre>
 
</pre>
 +
or you can use G notation for gigabyte
 +
<pre>
 +
vzctl set 123 --diskspace 20G:25G --save
 +
</pre>
 +
Here is 20 gigabyte barrier and 25 gigabyte limit diskspace saved. It can be checked by
 +
<pre>
 +
  # vzctl exec 123 df -h
 +
Filesystem            Size  Used Avail Use% Mounted on
 +
simfs                  20G  2.1G  18G  11% /
 +
tmpfs                443M    0  443M  0% /lib/init/rw
 +
tmpfs                443M    0  443M  0% /dev/shm
 +
</pre>
 
</li>
 
</li>
  
Line 99: Line 111:
 
</ol>
 
</ol>
  
 +
Applying a different diskinodes configuration is applied instantly and does not require a restart of the container.
 +
 +
You can also find the number of free inodes with
 +
 +
<pre>
 +
# vzctl exec 123 df -i
 +
</pre>
 
{{Note|shell does not support floating-point arithmetic, i.e. you can not use expressions like <code>$((&nbsp;220000*1.5&nbsp;))</code>. To use floating point, try <code>bc</code> instead, something like this: <code><nowiki>$(echo&nbsp;220000*1.5&nbsp;|&nbsp;bc)</nowiki></code>.}}
 
{{Note|shell does not support floating-point arithmetic, i.e. you can not use expressions like <code>$((&nbsp;220000*1.5&nbsp;))</code>. To use floating point, try <code>bc</code> instead, something like this: <code><nowiki>$(echo&nbsp;220000*1.5&nbsp;|&nbsp;bc)</nowiki></code>.}}
  
Line 123: Line 142:
 
Add the three units, 1000+2000+3000 = 6000
 
Add the three units, 1000+2000+3000 = 6000
  
101 gets 1000/6000 or 1/6th of the time. (16%)
+
* 101 gets 1000/6000 or 1/6th of the time. (16%)
102 gets 2000/6000 or 1/3rd of the time. (34%)
+
* 102 gets 2000/6000 or 1/3rd of the time. (34%)
103 gets 3000/6000 or 1/2 of the time.  (50%)
+
* 103 gets 3000/6000 or 1/2 of the time.  (50%)
 +
 
 +
To summarize: those units are proportional to each other. To say it more strict, to the sum of all CTs units, plus the host system, please don't forget that one. So indeed, units of 1 1 1 1 are the same as 200 200 200 200 or 8888 8888 8888 8888.
 +
 
 +
You may wonder why there's the tool vzcpucheck, which returns an absolute number called the "power of the node". The thing is, when you move a CT from one box to another, it could be problematic if you use different scales and different CPUs.
 +
 
 +
So vzcpucheck tries to work around that by inventing something called 'power of the node' which it gets from /proc/cpuinfo I guess (haven't checked it). If it shows a power of the node 10000 and you distribute that among all the CTs on the node, and then move one CT to another node which had cpuunits set in the same manner, that CT will have about the same CPU units it had on the old node.
  
 
=== cpulimit ===
 
=== cpulimit ===
Line 135: Line 160:
 
says that container 101 cannot ever have more than 25 percent of a CPU even if the CPU is idle for the other 75% of the time. The limit is calculated as a percentage of a single CPU, not as a percentage of the server's CPU resources as a whole. In other words, if you have more than one CPU, you can set a cpulimit > 100. In a quad-core server, setting cpulimit to 100 permits a container to consume one entire core (and not 100% of the server).
 
says that container 101 cannot ever have more than 25 percent of a CPU even if the CPU is idle for the other 75% of the time. The limit is calculated as a percentage of a single CPU, not as a percentage of the server's CPU resources as a whole. In other words, if you have more than one CPU, you can set a cpulimit > 100. In a quad-core server, setting cpulimit to 100 permits a container to consume one entire core (and not 100% of the server).
  
{{Warning|cpulimit is not yet implemented in kernels > 2.6.18 (i.e. development ones). Use stable kernel if you want this feature.}}
+
CPU limits are only available in rhel5-based and rhel6-based kernels and they behave a bit differently in them.
 +
 
 +
In the rhel5 kernel the limit has a container-wide meaning. That said if you have e.g. a container of 2 CPUS with the 100% cpulimit set, this container's usage of CPUs can be 100%/0% or 50%/50% or any other values, whose sum is 100%.
  
{{Stub}}
+
In the rhel6 kernel the applied limit is divided between onlince CPUs proportionally and a busy CPU cannot borrow time from an idle one. I.e. with a 2 CPUs container and 100% limit set the usage of each CPU cannot exceed 50% in any case.
  
 
[[Category: Troubleshooting]]
 
[[Category: Troubleshooting]]

Latest revision as of 19:34, 8 October 2013

Sometimes you see strange failures from some programs inside your container. In some cases it means one of the resources controlled by OpenVZ has hit the limit.

The first thing to do is to check the contents of the /proc/user_beancounters file in your container. The last column of output is the fail counter. Each time a resource hits the limit, the fail counter is incremented. So, if you see non-zero values in the failcnt column that means something is wrong.

There are two ways to fix the situation: reconfigure (in some cases recompile) the application, or change the resource management settings.

UBC parameters[edit]

Here is an example of current UBC values obtained from /proc/user_beancounters file in container 123:

# cat /proc/user_beancounters
Version: 2.5
       uid  resource           held    maxheld    barrier      limit    failcnt
       123: kmemsize         836919    1005343    2752512    2936012          0
            lockedpages           0          0         32         32          0
            privvmpages        4587       7289      49152      53575          0
            shmpages             39         39       8192       8192          0
            dummy                 0          0          0          0          0
            numproc              20         26         65         65          0
            physpages          2267       2399          0 2147483647          0
            vmguarpages           0          0       6144 2147483647          0
            oomguarpages       2267       2399       6144 2147483647          0
            numtcpsock            3          3         80         80          0
            numflock              3          4        100        110          0
            numpty                1          1         16         16          0
            numsiginfo            0          1        256        256          0
            tcpsndbuf             0          0     319488     524288          0
            tcprcvbuf             0          0     319488     524288          0
            othersockbuf       6684       7888     132096     336896          0
            dgramrcvbuf           0       8372     132096     132096          0
            numothersock          8         10         80         80          0
            dcachesize        87672      92168    1048576    1097728          0
            numfile             238        306       2048       2048          0
            dummy                 0          0          0          0          0
            dummy                 0          0          0          0          0
            dummy                 0          0          0          0          0
            numiptent            10         16        128        128          0

You can see if you hit the limit for some UBC parameters by analyzing the last column (named failcnt). It shows a number of failures for this counter, i.e. a number of times a parameter hit the limit. Usually what you need to do is to increase the parameter in question. But you need to do it carefully, and here is how.

  1. Get the current values for the parameter's barrier and limit. For example, we want to increase kmemsize values. From /proc/user_beancounters we see that kmemsize barrier is 2752512, and its limit is 2936012.
  2. Increase the values. Say, we want to double kmemsize. This is how it can be done using built-in bash arithmetics:
    # vzctl set 123 --kmemsize $((2752512*2)):$((2936012*2)) --save
    

    By using the --save flag, we indicate we want to apply the new setting to the running container and save it in the configuration file (from which the settings will be taken during next container start).

  3. Check the new configuration. Issue the following command:
    # vzcfgvalidate /etc/vz/conf/123.conf
    

    If something is wrong, you need to fix it as suggested by the utility.

For more in-depth explanation of different parameters, their meaning and how to set them properly, see setting UBC parameters.

Disk quota[edit]

To check if your container exceeded its disk quota, use the following commands (inside a container):

# df
Filesystem           1K-blocks      Used Available Use% Mounted on
simfs                  1048576    327664    720912  32% /
# df -i
Filesystem            Inodes   IUsed   IFree IUse% Mounted on
simfs                 200000   18857  181143   10% /

The first command shows disk space usage and the second command shows the inodes usage (you can roughly use the inodes count as a number of files/directories on your system).

If one of the commands shows a usage of 100% you have exceeded one of the disk quota limits.

You can increase the limit from the host system (CT0 aka VE0) only. This is how:

  1. Get the current values for disk quota:
    # vzquota stat 123
       resource          usage       softlimit      hardlimit    grace
      1k-blocks         327664         1048576        1153434
         inodes          18857          200000         220000
    
  2. To increase the disk space quota, use vzctl set --diskspace. For example, we want to increase it by a factor of 2:
    vzctl set 123 --diskspace $(( 1048576*2 )):$(( 1153434*2 )) --save
    

    or you can use G notation for gigabyte

    vzctl set 123 --diskspace 20G:25G --save
    

    Here is 20 gigabyte barrier and 25 gigabyte limit diskspace saved. It can be checked by

      # vzctl exec 123 df -h
    Filesystem            Size  Used Avail Use% Mounted on
    simfs                  20G  2.1G   18G  11% /
    tmpfs                 443M     0  443M   0% /lib/init/rw
    tmpfs                 443M     0  443M   0% /dev/shm
    
  3. To increase the disk inodes quota, use vzctl set --diskinodes. For example, we want to increase it by a factor of 3:
    vzctl set 123 --diskinodes $(( 200000*3 )):$(( 220000*3 )) --save
    

Applying a different diskinodes configuration is applied instantly and does not require a restart of the container.

You can also find the number of free inodes with

# vzctl exec 123 df -i
Yellowpin.svg Note: shell does not support floating-point arithmetic, i.e. you can not use expressions like $(( 220000*1.5 )). To use floating point, try bc instead, something like this: $(echo 220000*1.5 | bc).

CPU[edit]

There are two parameters controlling fair CPU scheduler in OpenVZ: cpuunits and cpulimit.

cpuunits[edit]

Cpuunits are set via

vzctl set 101 --cpuunits 1000 --save

For example. If you set a cpuunit for one container to a value and set a cpuunit on another container to a different value, the time allotted to each of the containers will be the ratio of the two units. Let's use a real example.

We did the following:

vzctl set 101 --cpuunits 1000 --save
vzctl set 102 --cpuunits 2000 --save
vzctl set 103 --cpuunits 3000 --save

If we started a CPU intensive application on each CT, then 103 would be given 3 times as much cpu time as 101 and 102 would get twice as much as 101, but some fraction of what 103 got. Here's how to determine what the real ratios are.

Add the three units, 1000+2000+3000 = 6000

  • 101 gets 1000/6000 or 1/6th of the time. (16%)
  • 102 gets 2000/6000 or 1/3rd of the time. (34%)
  • 103 gets 3000/6000 or 1/2 of the time. (50%)

To summarize: those units are proportional to each other. To say it more strict, to the sum of all CTs units, plus the host system, please don't forget that one. So indeed, units of 1 1 1 1 are the same as 200 200 200 200 or 8888 8888 8888 8888.

You may wonder why there's the tool vzcpucheck, which returns an absolute number called the "power of the node". The thing is, when you move a CT from one box to another, it could be problematic if you use different scales and different CPUs.

So vzcpucheck tries to work around that by inventing something called 'power of the node' which it gets from /proc/cpuinfo I guess (haven't checked it). If it shows a power of the node 10000 and you distribute that among all the CTs on the node, and then move one CT to another node which had cpuunits set in the same manner, that CT will have about the same CPU units it had on the old node.

cpulimit[edit]

The cpulimit parameter sets the absolute maximum limit for a container to a percent value. For instance:

vzctl set 101 --cpulimit 25 --save

says that container 101 cannot ever have more than 25 percent of a CPU even if the CPU is idle for the other 75% of the time. The limit is calculated as a percentage of a single CPU, not as a percentage of the server's CPU resources as a whole. In other words, if you have more than one CPU, you can set a cpulimit > 100. In a quad-core server, setting cpulimit to 100 permits a container to consume one entire core (and not 100% of the server).

CPU limits are only available in rhel5-based and rhel6-based kernels and they behave a bit differently in them.

In the rhel5 kernel the limit has a container-wide meaning. That said if you have e.g. a container of 2 CPUS with the 100% cpulimit set, this container's usage of CPUs can be 100%/0% or 50%/50% or any other values, whose sum is 100%.

In the rhel6 kernel the applied limit is divided between onlince CPUs proportionally and a busy CPU cannot borrow time from an idle one. I.e. with a 2 CPUs container and 100% limit set the usage of each CPU cannot exceed 50% in any case.