diff options
Diffstat (limited to 'docs/guides/configure/performance.md')
| -rw-r--r-- | docs/guides/configure/performance.md | 269 |
1 files changed, 108 insertions, 161 deletions
diff --git a/docs/guides/configure/performance.md b/docs/guides/configure/performance.md index 256d6e85..2e5e105f 100644 --- a/docs/guides/configure/performance.md +++ b/docs/guides/configure/performance.md @@ -1,110 +1,101 @@ -<!-- -title: How to optimize the Netdata Agent's performance -description: "While the Netdata Agent is designed to monitor a system with only 1% CPU, you can optimize its performance for low-resource systems." -image: /img/seo/guides/configure/performance.png -custom_edit_url: https://github.com/netdata/netdata/edit/master/docs/guides/configure/performance.md ---> - # How to optimize the Netdata Agent's performance We designed the Netdata Agent to be incredibly lightweight, even when it's collecting a few thousand dimensions every -second and visualizing that data into hundreds of charts. When properly configured for a production node, the Agent -itself should never use more than 1% of a single CPU core, roughly 50-100 MiB of RAM, and minimal disk I/O to collect, -store, and visualize all this data. - -We take this scalability seriously. We have one user [running -Netdata](https://github.com/netdata/netdata/issues/1323#issuecomment-266427841) on a system with 144 cores and 288 -threads. Despite collecting 100,000 metrics every second, the Agent still only uses 9% CPU utilization on a -single core. - -But not everyone has such powerful systems at their disposal. For example, you might run the Agent on a cloud VM with -only 512 MiB of RAM, or an IoT device like a [Raspberry Pi](https://github.com/netdata/netdata/blob/master/docs/guides/monitor/pi-hole-raspberry-pi.md). In these -cases, reducing Netdata's footprint beyond its already diminutive size can pay big dividends, giving your services more -horsepower while still monitoring the health and the performance of the node, OS, hardware, and applications. +second and visualizing that data into hundreds of charts. However, the default settings of the Netdata Agent are not +optimized for performance, but for a simple, standalone setup. We want the first install to give you something you can +run without any configuration. Most of the settings and options are enabled, since we want you to experience the full thing. -The default settings of the Netdata Agent are not optimized for performance, but for a simple standalone setup. We want -the first install to give you something you can run without any configuration. Most of the settings and options are -enabled, since we want you to experience the full thing. +By default, Netdata will automatically detect applications running on the node it is installed to start collecting metrics in +real-time, has health monitoring enabled to evaluate alerts and trains Machine Learning (ML) models for each metric, to detect anomalies. +This document describes the resources required for the various default capabilities and the strategies to optimize Netdata for production use. -## Prerequisites +## Summary of performance optimizations -- A node running the Netdata Agent. -- Familiarity with configuring the Netdata Agent with `edit-config`. +The following table summarizes the effect of each optimization on the CPU, RAM and Disk IO utilization in production. -If you're not familiar with how to configure the Netdata Agent, read our [node configuration -doc](https://github.com/netdata/netdata/blob/master/docs/configure/nodes.md) before continuing with this guide. This guide assumes familiarity with the Netdata config -directory, using `edit-config`, and the process of uncommenting/editing various settings in `netdata.conf` and other -configuration files. +Optimization | CPU | RAM | Disk IO +-- | -- | -- |-- +[Use streaming and replication](#use-streaming-and-replication) | :heavy_check_mark: | :heavy_check_mark: | :heavy_check_mark: +[Disable unneeded plugins or collectors](#disable-unneeded-plugins-or-collectors) | :heavy_check_mark: | :heavy_check_mark: | :heavy_check_mark: +[Reduce data collection frequency](#reduce-collection-frequency) | :heavy_check_mark: | | :heavy_check_mark: +[Change how long Netdata stores metrics](https://github.com/netdata/netdata/blob/master/docs/store/change-metrics-storage.md) | | :heavy_check_mark: | :heavy_check_mark: +[Use a different metric storage database](https://github.com/netdata/netdata/blob/master/database/README.md) | | :heavy_check_mark: | :heavy_check_mark: +[Disable machine learning](#disable-machine-learning) | :heavy_check_mark: | | +[Use a reverse proxy](#run-netdata-behind-a-proxy) | :heavy_check_mark: | | +[Disable/lower gzip compression for the agent dashboard](#disablelower-gzip-compression-for-the-dashboard) | :heavy_check_mark: | | -## What affects Netdata's performance? +## Resources required by a default Netdata installation Netdata's performance is primarily affected by **data collection/retention** and **clients accessing data**. -You can configure almost all aspects of data collection/retention, and certain aspects of clients accessing data. For -example, you can't control how many users might be viewing a local Agent dashboard, [viewing an -infrastructure](https://github.com/netdata/netdata/blob/master/docs/visualize/overview-infrastructure.md) in real-time with Netdata Cloud, or running [Metric -Correlations](https://github.com/netdata/netdata/blob/master/docs/cloud/insights/metric-correlations.md). +You can configure almost all aspects of data collection/retention, and certain aspects of clients accessing data. + +### CPU consumption + +Expect about: + - 1-3% of a single core for the netdata core + - 1-3% of a single core for the various collectors (e.g. go.d.plugin, apps.plugin) + - 5-10% of a single core, when ML training runs -The Netdata Agent runs with the lowest possible [process scheduling -policy](https://github.com/netdata/netdata/blob/master/daemon/README.md#netdata-process-scheduling-policy), which is `nice 19`, and uses the `idle` process scheduler. +Your experience may vary depending on the number of metrics collected, the collectors enabled and the specific environment they +run on, i.e. the work they have to do to collect these metrics. + +As a general rule, for modern hardware and VMs, the total CPU consumption of a standalone Netdata installation, including all its components, +should be below 5 - 15% of a single core. For example, on 8 core server it will use only 0.6% - 1.8% of a total CPU capacity, depending on +the CPU characteristics. + +The Netdata Agent runs with the lowest possible [process scheduling policy](https://github.com/netdata/netdata/blob/master/daemon/README.md#netdata-process-scheduling-policy), which is `nice 19`, and uses the `idle` process scheduler. Together, these settings ensure that the Agent only gets CPU resources when the node has CPU resources to space. If the node reaches 100% CPU utilization, the Agent is stopped first to ensure your applications get any available resources. -In addition, under heavy load, collectors that require disk I/O may stop and show gaps in charts. -Let's walk through the best ways to improve the Netdata Agent's performance. +To reduce CPU usage you can [disable machine learning](#disable-machine-learning), +[use streaming and replication](#use-streaming-and-replication), +[reduce the data collection frequency](#reduce-collection-frequency), [disable unneeded plugins or collectors](#disable-unneeded-plugins-or-collectors), [use a reverse proxy](#run-netdata-behind-a-proxy), and [disable/lower gzip compression for the agent dashboard](#disablelower-gzip-compression-for-the-dashboard). -## Reduce collection frequency +### Memory consumption -The fastest way to improve the Agent's resource utilization is to reduce how often it collects metrics. +The memory footprint of Netdata is mainly influenced by the number of metrics concurrently being collected. Expect about 150MB of RAM for a typical 64-bit server collecting about 2000 to 3000 metrics. -### Global +To estimate and control memory consumption, you can [disable unneeded plugins or collectors](#disable-unneeded-plugins-or-collectors), [change how long Netdata stores metrics](https://github.com/netdata/netdata/blob/master/docs/store/change-metrics-storage.md), or [use a different metric storage database](https://github.com/netdata/netdata/blob/master/database/README.md). -If you don't need per-second metrics, or if the Netdata Agent uses a lot of CPU even when no one is viewing that node's -dashboard, configure the Agent to collect metrics less often. -Open `netdata.conf` and edit the `update every` setting. The default is `1`, meaning that the Agent collects metrics -every second. +### Disk footprint and I/O -If you change this to `2`, Netdata enforces a minimum `update every` setting of 2 seconds, and collects metrics every -other second, which will effectively halve CPU utilization. Set this to `5` or `10` to collect metrics every 5 or 10 -seconds, respectively. +By default, Netdata should not use more than 1GB of disk space, most of which is dedicated for storing metric data and metadata. For typical installations collecting 2000 - 3000 metrics, this storage should provide a few days of high-resolution retention (per second), about a month of mid-resolution retention (per minute) and more than a year of low-resolution retention (per hour). -```conf -[global] - update every = 5 -``` +Netdata spreads I/O operations across time. For typical standalone installations there should be a few write operations every 5-10 seconds of a few kilobytes each, occasionally up to 1MB. In addition, under heavy load, collectors that require disk I/O may stop and show gaps in charts. -### Specific plugin or collector +To configure retention, you can [change how long Netdata stores metrics](https://github.com/netdata/netdata/blob/master/docs/store/change-metrics-storage.md). +To control disk I/O [use a different metric storage database](https://github.com/netdata/netdata/blob/master/database/README.md), avoid querying the +production system [using streaming and replication](#use-streaming-and-replication), [reduce the data collection frequency](#reduce-collection-frequency), and [disable unneeded plugins or collectors](#disable-unneeded-plugins-or-collectors). -Every collector and plugin has its own `update every` setting, which you can also change in the `go.d.conf`, -`python.d.conf`, or `charts.d.conf` files, or in individual collector configuration files. If the `update -every` for an individual collector is less than the global, the Netdata Agent uses the global setting. See the [enable -or configure a collector](https://github.com/netdata/netdata/blob/master/docs/collect/enable-configure.md) doc for details. +## Use streaming and replication -To reduce the frequency of an [internal -plugin/collector](https://github.com/netdata/netdata/blob/master/docs/collect/how-collectors-work.md#collector-architecture-and-terminology), open `netdata.conf` and -find the appropriate section. For example, to reduce the frequency of the `apps` plugin, which collects and visualizes -metrics on application resource utilization: +For all production environments, parent Netdata nodes outside the production infrastructure should be receiving all +collected data from children Netdata nodes running on the production infrastructure, using [streaming and replication](https://github.com/netdata/netdata/blob/master/docs/metrics-storage-management/enable-streaming.md). -```conf -[plugin:apps] - update every = 5 -``` +### Disable health checks on the child nodes -To [configure an individual collector](https://github.com/netdata/netdata/blob/master/docs/collect/enable-configure.md), open its specific configuration file with -`edit-config` and look for the `update_every` setting. For example, to reduce the frequency of the `nginx` collector, -run `sudo ./edit-config go.d/nginx.conf`: +When you set up streaming, we recommend you run your health checks on the parent. This saves resources on the children +and makes it easier to configure or disable alerts and agent notifications. + +The parents by default run health checks for each child, as long as the child is connected (the details are in `stream.conf`). +On the child nodes you should add to `netdata.conf` the following: ```conf -# [ GLOBAL ] -update_every: 10 +[health] + enabled = no ``` +### Use memory mode ram or save for the child nodes + +See [using a different metric storage database](https://github.com/netdata/netdata/blob/master/database/README.md). + ## Disable unneeded plugins or collectors If you know that you don't need an [entire plugin or a specific -collector](https://github.com/netdata/netdata/blob/master/docs/collect/how-collectors-work.md#collector-architecture-and-terminology), you can disable any of them. +collector](https://github.com/netdata/netdata/blob/master/collectors/README.md#collector-architecture-and-terminology), you can disable any of them. Keep in mind that if a plugin/collector has nothing to do, it simply shuts down and does not consume system resources. You will only improve the Agent's performance by disabling plugins/collectors that are actively collecting metrics. @@ -137,42 +128,60 @@ modules: fail2ban: no ``` -## Lower memory usage for metrics retention +## Reduce collection frequency + +The fastest way to improve the Agent's resource utilization is to reduce how often it collects metrics. + +### Global + +If you don't need per-second metrics, or if the Netdata Agent uses a lot of CPU even when no one is viewing that node's +dashboard, [configure the Agent](https://github.com/netdata/netdata/blob/master/docs/configure/nodes.md) to collect metrics less often. + +Open `netdata.conf` and edit the `update every` setting. The default is `1`, meaning that the Agent collects metrics +every second. + +If you change this to `2`, Netdata enforces a minimum `update every` setting of 2 seconds, and collects metrics every +other second, which will effectively halve CPU utilization. Set this to `5` or `10` to collect metrics every 5 or 10 +seconds, respectively. -Reduce the disk space that the [database engine](https://github.com/netdata/netdata/blob/master/database/engine/README.md) uses to retain metrics by editing -the `dbengine multihost disk space` option in `netdata.conf`. The default value is `256`, but can be set to a minimum of -`64`. By reducing the disk space allocation, Netdata also needs to store less metadata in the node's memory. +```conf +[global] + update every = 5 +``` -The `page cache size` option also directly impacts Netdata's memory usage, but has a minimum value of `32`. +### Specific plugin or collector -Reducing the value of `dbengine multihost disk space` does slim down Netdata's resource usage, but it also reduces how -long Netdata retains metrics. Find the right balance of performance and metrics retention by using the [dbengine -calculator](https://github.com/netdata/netdata/blob/master/docs/store/change-metrics-storage.md#calculate-the-system-resources-ram-disk-space-needed-to-store-metrics). +Every collector and plugin has its own `update every` setting, which you can also change in the `go.d.conf`, +`python.d.conf`, or `charts.d.conf` files, or in individual collector configuration files. If the `update +every` for an individual collector is less than the global, the Netdata Agent uses the global setting. See the [collectors configuration reference](https://github.com/netdata/netdata/blob/master/collectors/REFERENCE.md) for details. -All the settings are found in the `[global]` section of `netdata.conf`: +To reduce the frequency of an [internal +plugin/collector](https://github.com/netdata/netdata/blob/master/collectors/README.md#collector-architecture-and-terminology), open `netdata.conf` and +find the appropriate section. For example, to reduce the frequency of the `apps` plugin, which collects and visualizes +metrics on application resource utilization: ```conf -[db] - memory mode = dbengine - page cache size = 32 - dbengine multihost disk space = 256 +[plugin:apps] + update every = 5 ``` -To save even more memory, you can disable the dbengine and reduce retention to just 30 minutes, as shown below: +To [configure an individual collector](https://github.com/netdata/netdata/blob/master/collectors/REFERENCE.md#configure-a-collector), open its specific configuration file with +`edit-config` and look for the `update_every` setting. For example, to reduce the frequency of the `nginx` collector, +run `sudo ./edit-config go.d/nginx.conf`: ```conf -[db] - storage tiers = 1 - mode = alloc - retention = 1800 +# [ GLOBAL ] +update_every: 10 ``` -Metric retention is not important in certain use cases, such as: - - Data collection nodes stream collected metrics collected to a centralization point. - - Data collection nodes export their metrics to another time series DB, or are scraped by Prometheus - - Netdata installed only during incidents, to get richer information. -In such cases, you may not want to use the dbengine at all and instead opt for memory mode -`memory mode = alloc` or `memory mode = none`. +## Lower memory usage for metrics retention + +See how to [change how long Netdata stores metrics](https://github.com/netdata/netdata/blob/master/docs/store/change-metrics-storage.md). + +## Use a different metric storage database + +Consider [using a different metric storage database](https://github.com/netdata/netdata/blob/master/database/README.md) when running Netdata on IoT devices, +and for children in a parent-child set up based on [streaming and replication](https://github.com/netdata/netdata/blob/master/docs/metrics-storage-management/enable-streaming.md). ## Disable machine learning @@ -185,34 +194,12 @@ with the following: enabled = no ``` -## Run Netdata behind Nginx +## Run Netdata behind a proxy -A dedicated web server like Nginx provides far more robustness than the Agent's internal [web server](https://github.com/netdata/netdata/blob/master/web/README.md). +A dedicated web server like nginx provides more robustness than the Agent's internal [web server](https://github.com/netdata/netdata/blob/master/web/README.md). Nginx can handle more concurrent connections, reuse idle connections, and use fast gzip compression to reduce payloads. -For details on installing Nginx as a proxy for the local Agent dashboard, see our [Nginx -doc](https://github.com/netdata/netdata/blob/master/docs/Running-behind-nginx.md). - -After you complete Nginx setup according to the doc linked above, we recommend setting `keepalive` to `1024`, and using -gzip compression with the following options in the `location /` block: - -```conf - location / { - ... - gzip on; - gzip_proxied any; - gzip_types *; - } -``` - -Finally, edit `netdata.conf` with the following settings: - -```conf -[global] - bind socket to IP = 127.0.0.1 - disconnect idle web clients after seconds = 3600 - enable web responses gzip compression = no -``` +For details on installing another web server as a proxy for the local Agent dashboard, see [reverse proxies](https://github.com/netdata/netdata/blob/master/docs/category-overview-pages/reverse-proxies.md). ## Disable/lower gzip compression for the dashboard @@ -235,43 +222,3 @@ Or to lower the default compression level: gzip compression level = 1 ``` -## Disable logs - -If you installation is working correctly, and you're not actively auditing Netdata's logs, disable them in -`netdata.conf`. - -```conf -[logs] - debug log = none - error log = none - access log = none -``` - -## Disable health checks - -If you are streaming metrics to parent nodes, we recommend you run your health checks on the parent, for all the metrics collected -by the children nodes. This saves resources on the children and makes it easier to configure or disable alerts and agent notifications. - -The parents by default run health checks for each child, as long as it is connected (the details are in `stream.conf`). -On the child nodes you should add to `netdata.conf` the following: - -```conf -[health] - enabled = no -``` - -## What's next? - -We hope this guide helped you better understand how to optimize the performance of the Netdata Agent. - -Now that your Agent is running smoothly, we recommend you [secure your nodes](https://github.com/netdata/netdata/blob/master/docs/configure/nodes.md) if you haven't -already. - -Next, dive into some of Netdata's more complex features, such as configuring its health watchdog or exporting metrics to -an external time-series database. - -- [Interact with dashboards and charts](https://github.com/netdata/netdata/blob/master/docs/visualize/interact-dashboards-charts.md) -- [Configure health alarms](https://github.com/netdata/netdata/blob/master/docs/monitor/configure-alarms.md) -- [Export metrics to external time-series databases](https://github.com/netdata/netdata/blob/master/docs/export/external-databases.md) - -[](<>) |