Editing UBC systemwide configuration

{{UBC toc}}

The [[UBC consistency check]] article discussed validation of [[UBC]] configuration for a single container. This article discusses checks that the UBC configuration of the
'''whole''' system is '''valid'''.

Configurations where resources allowed for containers exceed system
capacity<ref>More precisely, configurations with excessive
overcommitment, as explained below.</ref> are not valid
and dangerous from stability point of view.
They may result in abnormal termination of the applications,
bad responsiveness of the system and, sometimes, system hangs.
Whereas the configuration validation discussed in [[UBC consistency check]]
addressed application functionality,
the validation considered in this section is aimed at security and stability
of the whole system.

The best way to make sure that the configuration of the whole system
is valid is to run periodic automatic checks, based on the formulae described
below. {{Man|vzmemcheck|8}} utility can be helpful in calculations.

== Resource utilization and commitment level ==

Several resources of the whole system (such as RAM) are discussed below
in terms of '''utilization''' and '''commitment level'''.

'''Utilization''' shows the amount of resources consumed by all
containers at the given time.
In general, low utilization values mean that the system is under-utilized.
Often, it means that the system is capable of supporting more containers
if the existing containers continue to maintain
the same load and resource consumption level.
High utilization values (in general, more than <code>1</code>)
mean the the system is overloaded and the service level
of the containers is degraded.

'''Commitment level''' shows how much resources are “promised” to the existing
containers. Low commitment levels mean that the system
is capable of supporting more containers.
Commitment levels more than <code>1</code> mean that the
containers are promised more resources
than the system has, and in this case the system is said to be
overcommitted.
If the system runs a lot of containers,
it is usually acceptable to have some overcommitment, because it is unlikely that all containers will request resources at the same time.
However, higher commitment levels (as discussed below for each resource
individually) will cause containerss to experience failures to
allocate and use the resources promised to them.

== “Low memory” (x86_32 specific) ==
Because of specifics of architecture of Intel's x86 processors, the RAM of the
computer can't be used uniformly.
The most important memory area is so called “low memory”, a part of memory
residing at lower addresses and directly accessible by the kernel.
For current Linux kernels, the size of low memory area is 832MB
(or, if the computer has less RAM than 832MB, the size of the RAM).

64-bit kernels (i.e x86_64, ia64 and others) can access all memory directly and are not used "low memory" area.
=== Utilization ===
The lower bound estimation of low memory utilization is

<math>
\frac{\displaystyle\sum_{all\ containers}
        (kmemsize_{cur}+allsocketbuf_{cur})}
    {0.4\cdot\min(RAM\ size, {\rm 832MB})}\rm,
</math>

where

<math>
allsocketbuf=tcprcvbuf+tcpsndbuf+dgramrcvbuf+othersockbuf\rm.
</math>

Utilization of low memory below <math>1</math> is normal.
Utilization above <math>1</math> is not safe, and utilization above <math>2</math>
is dangerous is very likely to cause bad system responsiveness,
application stalls for seconds or more and termination of some applications.

=== Commitment level ===
The commitment level can be computed as

<math>
\frac{\displaystyle\sum_{all\ containers}
        (kmemsize_{lim}+allsocketbuf_{lim})}
    {0.4\cdot\min(RAM\ size, {\rm 832MB})}\rm.
</math>

Commitment levels below <math>1</math> are normal.
Levels between <math>1</math> and <math>1.2</math> are usually acceptable for systems with about
100 containers.
Systems with more containers may have the commitment level
increased, to about <math>1.5</math>—<math>2</math> for 400 containers.
Higher commitment levels for this resource are not recommended,
because the consequences of exceeding the low memory capacities
are severe and affect the whole system and all the containers.

== Total RAM ==
This subsection discusses usage of the whole RAM and its utilization.
Usage of swap space and the sum of used RAM and swap space are discussed
below in subsection [[{{PAGENAME}}#Memory and swap space|Memory and swap space]].

Current version of OpenVZ can't guarantee availability of certain
amount of memory (in opposite to the sum of memory and swap space),
so the commitment level is not applicable to the total RAM.
Such guarantees will be implemented in future versions.

=== Utilization ===

The amount of RAM consumed by all containers can be computed as

<math>
\sum_{all\ containers}
    (physpages_{cur}\cdot4096+kmemsize_{cur}+allsocketbuf_{cur})\rm.
</math>

The difference between memory usage shown by <code>free(1)</code>
or <code>/proc/meminfo</code> and the total amount of RAM consumed by
containers is the memory used by system daemons
and different caches.

The memory utilization can be computed as

<math>
\frac{\displaystyle\sum_{all\ containers}
        (physpages_{cur}\cdot4096+
         kmemsize_{cur}+allsocketbuf_{cur})}
    {RAM\ size}\rm.
</math>

Utilization levels from <math>0.8</math> to <math>1</math> are normal.
Lower utilization means that the system is under-utilized, and,
if other system resources and their commitment levels permit,
the system can host more containers.
By the nature of the accounting of <code>physpages</code> and other parameters,
total RAM utilization can't be bigger than <math>1</math>.

== Memory and swap space ==
The main resource of the computer determining the amount of memory
applications can use is the sum of RAM and swap sizes.
If the total size of the used memory exceeds the RAM size,
Linux kernel moves some data to swap and and loads it back when the
application needs it.
More frequently used data tends to stay in RAM, less frequently used
data spends more time in swap.

Swap-in and swap-out activity reduces the system performance to some extent.
However, if this activity is not excessive, the performance decrease is not
very noticeable.  On the other hand, the benefits of using swap space are
quite big, allowing to increase the number of containers in the
system about 2 times.

Swap space is essential for handling system load bursts.
System with enough swap just slows down at high load bursts,
whereas the system without swap reacts to high load bursts by refusing
memory allocations (causing applications to refuse to accept clients or
terminate) and by direct killing of some applications.
Additionally, the presence of swap space helps the system to better
balance memory and move data between “low memory” and the rest of the RAM.

In all OpenVZ installations it is strongly recommended
to have swap of size not less than the RAM size.

Also, it is not recommended to create swap space of the size
of more than 4 times RAM size because of performance degradation
related to swap-in and swap-out activity.
That is, the system should be configured so that
<math>RAM\ size \le swap\ size \le 4\cdot RAM\ size\rm.</math>
The optimal configuration is when swap size is twice more than the RAM size.

=== Utilization ===
<math>
\frac{\displaystyle\sum_{all\ containers}
        (oomguarpages_{cur}\cdot4096+
         kmemsize_{cur}+allsocketbuf_{cur})}
    {RAM\ size + swap\ size}\rm.
</math>

The normal utilization of memory plus swap ranges between
<math>\frac{RAM\ size}{RAM\ size + swap\ size}</math>
and
<math>\frac{RAM\ size + \frac12swap\ size}{RAM\ size + swap\ size}.</math>

Lower utilization means that the system memory is under-utilized
at the moment of checking the utilization.
Higher utilization is likely to cause gradual performance degradation because
of swap-in and swap-out activity and is a sign of overloading of the system.

=== Commitment level ===
<math>
\frac{\displaystyle\sum_{all\ containers}
        (oomguarpages_{bar}\cdot4096+
         kmemsize_{lim}+allsocketbuf_{lim})}
    {RAM\ size + swap\ size}\rm.
</math>

The normal commitment level is about <math>0.8</math>—<math>1</math>.

Commitment levels more than <math>1</math> means that the containers are
guaranteed more memory than the system has.
Such overcommitment is strongly not recommended, because in that case
if all the memory is consumed, random applications, including the ones
belonging to the host system, may be killed and the system may become
inaccessible by <code>ssh(1)</code> and lose other important functionality.

It is better to guarantee containers less and have less commitment
levels than to accidentally overcommit the system by memory plus swap.
If the system has spare memory and swap, containers will
transparently be able to use the memory and swap above their guarantees.
Guarantees given to containers should not be big, and it is normal
if memory and swap usage for some containers stays above their
guarantee.
It is also normal to give guarantees only to containers with
preferred service.
But administrators should not guarantee container more
than the system actually has.

== Allocated memory ==

This subsection considers standard memory allocations made by
applications in the container.
The allocations for each container are controlled by
two parameters: <code>vmguarpages</code> and <code>privvmpages</code>,
discussed in sections FIXME.

Allocated memory is a more “virtual” system resource than the RAM or RAM
plus swap size.
Applications can allocate memory but start to use it only later, and
the amount of system's free memory will decrease only at the moment of the
use.
The sum of allocated memory size of all containers is an estimation
of how much physical memory will be used when (and if) all applications
claim the allocated memory.

=== Utilization ===
<math>
\frac{\displaystyle\sum_{all\ containers}
        (privvmpages_{cur}\cdot4096+
         kmemsize_{cur}+allsocketbuf_{cur})}
    {RAM\ size + swap\ size}\rm.
</math>

This utilization level is the ratio of the amount of allocated memory
to the capacity of the system.

Low utilization level means that the system can support more
containers, if other resources permit.
High utilization levels may, but doesn't necessarily mean that
the system is overloaded.
As it was explaned above, not all applications use all the allocated memory,
so this utilization level may exceed <math>1</math>.

Computing this utilization level is useful for comparing
it with the commitment level and the level of memory allocation restrictions,
discussed below, to configure memory allocation restrictions
for the container.

=== Commitment level ===
Allocation guarantee commitment level

<math>
\frac{\displaystyle\sum_{all\ containers}
        (vmguarpages_{bar}\cdot4096+
         kmemsize_{lim}+allsocketbuf_{lim})}
    {RAM\ size + swap\ size}
</math>

is the ratio of the memory space guaranteed to be available for
allocations to the capacity of the system.
Similarly to the commitment level of memory plus swap space (as discussed
in subsection [[{{PAGENAME}}#Memory and swap space|Memory and swap space]]),
this level should be kept below <math>1</math>.
If the level is above <math>1</math>, it significantly increases the chances of
applications to be killed instead of be notified in case the system
experiences a memory shortage.

It's better to provide lower guarantees than accidently guarantee more than
the system has, because containers are allowed to allocated memory
above their guarantee if the system is not tight on memory.
It is also normal to give guarantees only to containers with
preferred service.

== Limiting memory allocations ==
In addition to providing allocation guarantees, it is possible
to impose restrictions on the amount of memory allocated by
containers.

If a system has multiple containers,
it is important to make sure that for each container

<math>privvmpages_{lim}\cdot4096 \le 0.6\cdot {RAM\ size}\rm.</math>

If this condition is not satisfied, a single container may easily
cause an excessive swap-out and very bad performance of the whole system.
Usually, for each container <code>privvmpages</code> limitations are
to values much less than the size of the RAM.

The resource control parameters should be configured in a way,
so that in case of memory shortage applications are given chance
to notice the shortage and exit gracefully, instead of being terminated
by the kernel.
For this purpose, it is recommended to maintain reasonable
total level of memory allocation restrictions, computed as

<math>
\frac{\displaystyle\sum_{all\ containers}
        (privvmpages_{lim}\cdot4096+
         kmemsize_{lim}+allsocketbuf_{lim})}
    {RAM\ size + swap\ size}\rm.
</math>

This number shows how much memory applications are allowed to allocate
in comparison with the capacity of the system.

In practice, a lot of applications do not use the memory very efficiently
and, sometimes, allocated memory will never be used later.
For example, Apache Web server at start time it allocates about
20–30%% more memory that it will ever use.
Some multi-threaded applications are especially bad at using their memory,
and their rate of allocated to used memory may happen to be 1000%.

The bigger the level of memory allocation restrictions,
the more chances are that applications will be killed instead
of getting an error on next memory allocation in case
the system experiences memory shortage.
The levels ranging in <math>1.5</math>–<math>4</math> can be considered acceptable.
Administrators can find experimentally the optimal setting for their load,
basing on the frequency of messages “Out of Memory: killing process”
in system logs, saved by <code>klogd(8)</code> and <code>syslogd(8)</code>.
However, for stability-critical applications, it's better to keep
the level not exceeding <math>1</math>.

<references/>