diff options
author | Daniel Baumann <daniel.baumann@progress-linux.org> | 2019-07-08 20:14:42 +0000 |
---|---|---|
committer | Daniel Baumann <daniel.baumann@progress-linux.org> | 2019-07-08 20:14:42 +0000 |
commit | 4f88e1a9be89a257fd6ed3045703db6e900027ee (patch) | |
tree | 518eb3c3aa1dce9ea281d02e0fd3cc01a9e7913f /docs/Performance.md | |
parent | Adding upstream version 1.15.0. (diff) | |
download | netdata-4f88e1a9be89a257fd6ed3045703db6e900027ee.tar.xz netdata-4f88e1a9be89a257fd6ed3045703db6e900027ee.zip |
Adding upstream version 1.16.0.upstream/1.16.0
Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>
Diffstat (limited to '')
-rw-r--r-- | docs/Performance.md | 52 |
1 files changed, 24 insertions, 28 deletions
diff --git a/docs/Performance.md b/docs/Performance.md index b08549f11..fbc6d5761 100644 --- a/docs/Performance.md +++ b/docs/Performance.md @@ -1,6 +1,6 @@ # Performance -netdata performance is affected by: +Netdata performance is affected by: **Data collection** - the number of charts for which data are collected @@ -19,11 +19,11 @@ You can control all the above. ## Netdata Daemon -For most server systems, with a few hundred charts and a few thousand dimensions, the netdata daemon, without any web clients accessing it, should not use more than 1% of a single core. +For most server systems, with a few hundred charts and a few thousand dimensions, the Netdata daemon, without any web clients accessing it, should not use more than 1% of a single core. -To prove netdata scalability, check issue [#1323](https://github.com/netdata/netdata/issues/1323#issuecomment-265501668) where netdata collects 95.000 metrics per second, with 12% CPU utilization of a single core! +To prove Netdata scalability, check issue [#1323](https://github.com/netdata/netdata/issues/1323#issuecomment-265501668) where Netdata collects 95.000 metrics per second, with 12% CPU utilization of a single core! -In embedded systems, if the netdata daemon is using a lot of CPU without any web clients accessing it, you should lower the data collection frequency. To set the data collection frequency, edit `/etc/netdata/netdata.conf` and set `update_every` to a higher number (this is the frequency in seconds data are collected for all charts: higher number of seconds = lower frequency, the default is 1 for per second data collection). You can also set this frequency per module or chart. Check the [daemon configuration](../daemon/config) for plugins and charts. For specific modules, the configuration needs to be changed in: +In embedded systems, if the Netdata daemon is using a lot of CPU without any web clients accessing it, you should lower the data collection frequency. To set the data collection frequency, edit `/etc/netdata/netdata.conf` and set `update_every` to a higher number (this is the frequency in seconds data are collected for all charts: higher number of seconds = lower frequency, the default is 1 for per second data collection). You can also set this frequency per module or chart. Check the [daemon configuration](../daemon/config) for plugins and charts. For specific modules, the configuration needs to be changed in: - `python.d.conf` for [python](../collectors/python.d.plugin/#pythondplugin) - `node.d.conf` for [nodejs](../collectors/node.d.plugin/#nodedplugin) - `charts.d.conf` for [bash](../collectors/charts.d.plugin/#chartsdplugin) @@ -34,24 +34,24 @@ If a plugin is using a lot of CPU, you should lower its update frequency, or if ## CPU consumption when web clients are accessing dashboards -Netdata is very efficient when servicing web clients. On most server platforms, netdata should be able to serve **1800 web client requests per second per core** for auto-refreshing charts. +Netdata is very efficient when servicing web clients. On most server platforms, Netdata should be able to serve **1800 web client requests per second per core** for auto-refreshing charts. Normally, each user connected will request less than 10 chart refreshes per second (the page may have hundreds of charts, but only the visible are refreshed). So you can expect 180 users per CPU core accessing dashboards before having any delays. Netdata runs with the lowest possible process priority, so even if 1000 users are accessing dashboards, it should not influence your applications. CPU utilization will reach 100%, but your applications should get all the CPU they need. -To lower the CPU utilization of netdata when clients are accessing the dashboard, set `web compression level = 1`, or disable web compression completely by setting `enable web responses gzip compression = no`. Both settings are in the `[web]` section. +To lower the CPU utilization of Netdata when clients are accessing the dashboard, set `web compression level = 1`, or disable web compression completely by setting `enable web responses gzip compression = no`. Both settings are in the `[web]` section. ## Monitoring a heavy loaded system -Netdata, while running, does not depend on disk I/O (apart its log files and `access.log` is written with buffering enabled and can be disabled). Some plugins that need disk may stop and show gaps during heavy system load, but the netdata daemon itself should be able to work and collect values from `/proc` and `/sys` and serve web clients accessing it. +Netdata, while running, does not depend on disk I/O (apart its log files and `access.log` is written with buffering enabled and can be disabled). Some plugins that need disk may stop and show gaps during heavy system load, but the Netdata daemon itself should be able to work and collect values from `/proc` and `/sys` and serve web clients accessing it. -Keep in mind that netdata saves its database when it exits and loads it back when restarted. While it is running though, its DB is only stored in RAM and no I/O takes place for it. +Keep in mind that Netdata saves its database when it exits and loads it back when restarted. While it is running though, its DB is only stored in RAM and no I/O takes place for it. ## Netdata process priority -By default, netdata runs with the `idle` process scheduler, which assigns CPU resources to netdata, only when the system has such resources to spare. +By default, Netdata runs with the `idle` process scheduler, which assigns CPU resources to Netdata, only when the system has such resources to spare. The following `netdata.conf` settings control this: @@ -62,15 +62,15 @@ The following `netdata.conf` settings control this: process nice level = 19 ``` -The policies supported by netdata are `idle` (the netdata default), `other` (also as `nice`), `batch`, `rr`, `fifo`. netdata also recognizes `keep` and `none` to keep the current settings without changing them. +The policies supported by Netdata are `idle` (the Netdata default), `other` (also as `nice`), `batch`, `rr`, `fifo`. Netdata also recognizes `keep` and `none` to keep the current settings without changing them. -For `other`, `nice` and `batch`, the setting `process nice level = 19` is activated to configure the nice level of netdata. Nice gets values -20 (highest) to 19 (lowest). +For `other`, `nice` and `batch`, the setting `process nice level = 19` is activated to configure the nice level of Netdata. Nice gets values -20 (highest) to 19 (lowest). For `rr` and `fifo`, the setting `process scheduling priority = 0` is activated to configure the priority of the relative scheduling policy. Priority gets values 1 (lowest) to 99 (highest). For the details of each scheduler, see `man sched_setscheduler` and `man sched`. -When netdata is running under systemd, it can only lower its priority (the default is `other` with `nice level = 0`). If you want to make netdata to get more CPU than that, you will need to set in `netdata.conf`: +When Netdata is running under systemd, it can only lower its priority (the default is `other` with `nice level = 0`). If you want to make Netdata to get more CPU than that, you will need to set in `netdata.conf`: ``` [global] @@ -85,17 +85,17 @@ CPUSchedulingPriority=99 Nice=-10 ``` -## Running netdata in embedded devices +## Running Netdata in embedded devices Embedded devices usually have very limited CPU resources available, and in most cases, just a single core. -> keep in mind that netdata on RPi 2 and 3 does not require any tuning. The default settings will be good. The following tunables apply only when running netdata on RPi 1 or other very weak IoT devices. +> keep in mind that Netdata on RPi 2 and 3 does not require any tuning. The default settings will be good. The following tunables apply only when running Netdata on RPi 1 or other very weak IoT devices. We suggest to do the following: ### 1. Disable External plugins -External plugins can consume more system resources than the netdata server. Disable the ones you don't need. If you need them, increase their `update every` value (again in `/etc/netdata/netdata.conf`), so that they do not run that frequently. +External plugins can consume more system resources than the Netdata server. Disable the ones you don't need. If you need them, increase their `update every` value (again in `/etc/netdata/netdata.conf`), so that they do not run that frequently. Edit `/etc/netdata/netdata.conf`, find the `[plugins]` section: @@ -126,8 +126,8 @@ plugin|description `idlejitter`|internal plugin (written in C) that attempts show if the systems starved for CPU. Disabling it will eliminate a thread. `cgroups`|monitoring linux containers. Most probably you are not going to need it. This will also eliminate another thread. `checks`|a debugging plugin, which is disabled by default. -`apps`|a plugin that monitors system processes. It is very complex and heavy (consumes twice the CPU resources of the netdata daemon), so if you don't need to monitor the process tree, you can disable it. -`charts.d`|BASH plugins (squid, nginx, mysql, etc). This is a heavy plugin, that consumes twice the CPU resources of the netdata daemon. +`apps`|a plugin that monitors system processes. It is very complex and heavy (consumes twice the CPU resources of the Netdata daemon), so if you don't need to monitor the process tree, you can disable it. +`charts.d`|BASH plugins (squid, nginx, mysql, etc). This is a heavy plugin, that consumes twice the CPU resources of the Netdata daemon. `node.d`|node.js plugin, currently used for SNMP data collection and monitoring named (the name server). `python.d`|has many modules and can use over 20MB of memory. @@ -137,7 +137,7 @@ For most IoT devices, you can disable all plugins except `proc`. For `proc` ther ### 2. Disable internal plugins -In this section you can select which modules of the `proc` plugin you need. All these are run in a single thread, one after another. Still, each one needs some RAM and consumes some CPU cycles. With all the modules enabled, the `proc` plugin adds ~9 MiB on top of the 5 MiB required by the netdata daemon. +In this section you can select which modules of the `proc` plugin you need. All these are run in a single thread, one after another. Still, each one needs some RAM and consumes some CPU cycles. With all the modules enabled, the `proc` plugin adds ~9 MiB on top of the 5 MiB required by the Netdata daemon. ``` [plugin:proc] @@ -150,7 +150,7 @@ Refer to the [proc.plugins documentation](../collectors/proc.plugin/) for the li ### 3. Lower internal plugin update frequency -If netdata is still using a lot of CPU, lower its update frequency. Going from per second updates, to once every 2 seconds updates, will cut the CPU resources of all netdata programs **in half**, and you will still have very frequent updates. +If Netdata is still using a lot of CPU, lower its update frequency. Going from per second updates, to once every 2 seconds updates, will cut the CPU resources of all Netdata programs **in half**, and you will still have very frequent updates. If the CPU of the embedded device is too weak, try setting even lower update frequency. Experiment with `update every = 5` or `update every = 10` (higher number = lower frequency) in `netdata.conf`, until you get acceptable results. @@ -172,18 +172,14 @@ Normally, you will not need them. To disable them, set: ``` ### 5. Set memory mode to RAM -Setting the memory mode to `ram` will disable loading and saving the round robin database. This will not affect anything while running netdata, but it might be required if you have very limited storage available. +Setting the memory mode to `ram` will disable loading and saving the round robin database. This will not affect anything while running Netdata, but it might be required if you have very limited storage available. ``` [global] memory mode = ram ``` -### 6. Use the single threaded web server - -Normally, netdata spawns a thread for each web client. This allows netdata to utilize all the available cores for servicing chart refreshes. You can however disable this feature and serve all charts one after another, using a single thread / core. This will might lower the CPU pressure on the embedded device. To enable the single threaded web server, edit `/etc/netdata/netdata.conf` and set `mode = single-threaded` in the `[web]` section. - -### 7. Lower memory requirements +### 6. Lower memory requirements You can set the default size of the round robin database for all charts, using: @@ -197,9 +193,9 @@ The units for history is `[global].update every` seconds. So if `[global].update Check also [Database](../database) for directions on calculating the size of the round robin database. -### 8. Disable gzip compression of responses +### 7. Disable gzip compression of responses -Gzip compression of the web responses is using more CPU that the rest of netdata. You can lower the compression level or disable gzip compression completely. You can disable it, like this: +Gzip compression of the web responses is using more CPU that the rest of Netdata. You can lower the compression level or disable gzip compression completely. You can disable it, like this: ``` [web] @@ -214,7 +210,7 @@ To lower the compression level, do this: gzip compression level = 1 ``` -Finally, if no web server is installed on your device, you can use port tcp/80 for netdata: +Finally, if no web server is installed on your device, you can use port tcp/80 for Netdata: ``` [web] |