| Latest revision |
Your text |
| Line 1: |
Line 1: |
| − | <translate>
| + | This page describes how to do correct performance measurements on OpenVZ system. |
| − | <!--T:1-->
| |
| − | This page describes how to improve the performance of an OpenVZ system. | |
| | | | |
| − | == HW node environment tuning == <!--T:2--> | + | = Test conditions = |
| | | | |
| − | === Disable unnecessary services === <!--T:3--> | + | == kernels with different versions == |
| | + | * If you want to compare performance of the kernel on different hosts, or if you want to measure OpenVZ performance overhead, it's strongly recommended to compare the same kernel version with similar .config file and maybe with the same linux distribution. |
| | | | |
| − | <!--T:4-->
| + | * If you compare kernels with different versions, please check all .config options that are differ, especially _DEBUG_ options. For example, on unixbench pipe throughput test on 2.6.18 kernel, disabled CONFIG_DEBUG_HIGHMEM option will increase performance up to <font color=red>20%</font> |
| − | Disable all default services that you do not need to use and then reboot your host.
| |
| | | | |
| − | <!--T:5-->
| + | == running services == |
| − | For example, the <code>audit</code> daemon can significantly decrease performance of linux kernel system calls (up to ~<font color=red>20%</font>) even if you do not use any audit rules, or even if you just stopped this service without host reboot!
| + | * Before performing test measurement, you should disable '''all''' default services in your runlevel and then '''reboot''' your host. |
| | | | |
| − | <!--T:6-->
| + | It is not enough to just stop services, because some services, like <code>audit</code> will affect performance even if the daemon is already stopped. For some unixbench tests it can be <font color=red>~20%</font> overhead if <code>auditd</code> was started once on a host before reboot. |
| − | To setup default services, use <code>chkconfig</code> or <code>ntsysv</code> in RedHat, or <code>rc-update</code> in Gentoo, <code>update-rc.d</code> on Debian
| |
| | | | |
| − | === Shell scripts performance improvement === <!--T:7-->
| + | On RedHat distributions use <code>chkconfig</code> or <code>ntsysv</code> utility to disable default services. (<code>rc-update</code> in Gentoo, <code>update-rc.dv</code> for Debian) |
| | | | |
| − | <!--T:8-->
| + | == filesystem tests == |
| − | To improve performance of small shell scripts, which spends a lot of time starting the shell binary itself (like the shell scripts test from the [http://www.tux.org/pub/tux/niemi/unixbench/ unixbench] package), you can set your <code>LANG</code> environment variable to <code>"C"</code>.
| + | * If you perform filesystem tests, please keep in mind filesystem type, block size, mount options and so on. |
| | | | |
| − | <!--T:9-->
| + | For example, ext3 filesystem performance highly depends on journal type and mount options. |
| − | To see current settings, type
| |
| | | | |
| − | <!--T:10-->
| + | * Also please always note/report IO-scheduler type. Different IO-schedulers can highly affect your tests results (up to <font color=red>30%</font>). |
| − | # locale
| |
| | | | |
| − | <!--T:11-->
| + | If your kernel support different IO-schedulers, you can get/set the type here: |
| − | If you want to change it only for the current shell session, do:
| |
| | | | |
| − | <!--T:12--> | + | # cat /sys/block/hda/queue/scheduler |
| − | # export LANG=C | + | noop anticipatory deadline [cfq] |
| | + | # echo noop > /sys/block/hda/queue/scheduler |
| | | | |
| − | <!--T:13-->
| + | == network isolation == |
| − | If you want to change the default value, modify the <code>/etc/sysconfig/i18n</code> file.
| + | * You should disable local network/internet connection if your tests doesn't require it. |
| | | | |
| − | <!--T:14-->
| + | == CPU distribution inside VE on SMP hosts == |
| − | If your default <code>LANG</code> environment variable was set to something like <code>en_US.UTF-8</code>, you can reduce shell (bash) startup time up to ~<font color=red>15%</font> with <code>LANG=C</code>. | + | * If the number of VE's in your host is more than CPUs number, and there are many tasks/tests running inside each VE, and that tasks are scheduled quite often, it's better to give just one CPU for each VE. In this case the VirtualCPU-scheduler performance overhead can be significantly decreased, and performance can increase up to <font color=red>100%</font>! |
| | | | |
| − | == Container tuning == <!--T:15-->
| + | To set the number of CPUs available inside VE use: |
| | | | |
| − | === CPU distribution inside container on SMP hosts === <!--T:16-->
| + | # vzctl set $VEID --cpus N |
| | | | |
| − | <!--T:17-->
| + | == network performance == |
| − | If the total number of containers in your host is more than CPUs number, and there are many '''threads''' running inside each container it is better to give just a single VCPU to each container.
| + | * please do not use file transferring utilities to test the network performance, because the bottleneck of these tests is usually file system performance - not TCP/IP stack |
| − | In this case thread memory locality will significantly reduce overhead on SMP memory coherence and overall performance can be increased up to ~<font color=red>50-100%</font>!
| |
| | | | |
| − | <!--T:18-->
| + | == network checksumming == |
| − | To set the number of CPUs available inside a container, use:
| + | '''TODO''' |
| − | | |
| − | <!--T:19-->
| |
| − | # vzctl set $CTID --cpus N
| |
| − | | |
| − | === Network checksumming === <!--T:20--> | |
| − | | |
| − | <!--T:21-->
| |
| − | RHEL 5 based kernel supports IP checksum offload.
| |
| − | If network ethernet cards in your host support IP checksum offload then you can switch this feature on also for the virtual network devices (venet, veth).
| |
| − | | |
| − | <!--T:22-->
| |
| − | To check current offload setting for the hardware ethernet card (eth0, for instance) type
| |
| − | | |
| − | <!--T:23-->
| |
| − | # ethtool -k eth0
| |
| − |
| |
| − | Make sure that tx/rx features are switched on.
| |
| − | | |
| − | <!--T:24-->
| |
| − | To see current offload settings for the venet0 device, type
| |
| − | | |
| − | <!--T:25-->
| |
| − | # ethtool -k venet0
| |
| − | | |
| − | <!--T:26-->
| |
| − | To set offload settings on for the venet0 device, type
| |
| − | | |
| − | <!--T:27-->
| |
| − | # ethtool -K venet0 tx on sg on
| |
| − | | |
| − | <!--T:28-->
| |
| − | Note, that 'tx on/off' enables/disables both tx and rx checksumming features for the all venet devices for all containers and HN.
| |
| − | | |
| − | <!--T:29-->
| |
| − | The same applies to the veth device except that 'tx on/off' enables/disables tx and rx checksumming features for only given virtual ethernet device in HN and corresponding container.
| |
| − | | |
| − | === Shell scripts performance improvement === <!--T:30-->
| |
| − | | |
| − | <!--T:31-->
| |
| − | Please note, that on container creation the default <code>LANG</code> value will be the same as in the HW node. So you can tune it in node (see [[#Shell scripts performance improvement]] above), or set it in container the same way.
| |
| − | | |
| − | <!--T:32-->
| |
| − | The second important thing is the locale cache. On <code>rpm</code> based distributions, usually it is created by the <code>glibc-common-XXX.rpm</code> post install script and it can be up to 50 MBytes on some distributions. So on some container templates it can be missed to save disk space. But you can always create it inside container later by the following command (you must be the root user):
| |
| − | | |
| − | <!--T:33-->
| |
| − | # build-locale-archive
| |
| − | | |
| − | <!--T:34-->
| |
| − | And again, in some cases shell (bash) startup time can be reduced up to ~<font color=red>15%</font>.
| |
| − | </translate>
| |
| − | | |
| − | [[Category: HOWTO]]
| |
| − | [[Category: Troubleshooting]]
| |