New: temperature, fans, power supplies and SSD health per server. View hardware health

Try it free
Server monitor web-01 in the portal: header with 21% CPU utilization, 58% RAM utilization, 41% disk and uptime, below the CPU and RAM history over the day with a load peak during the night.

PeakNode monitors your servers around the clock: CPU, RAM, every drive and the network. If a disk fills up, RAM runs short or a server stops reporting, you hear about it by email, Teams or Slack. All that runs on the server is an agent, a script that sends readings outbound.

Screenshot from the portal with sample data (German interface).

What is Server monitoring?

Server monitoring with PeakNode means: an agent, a script on the server, measures CPU utilization, RAM, swap, the usage of every drive and network throughput, and sends the values to the portal at the configured interval. The portal stores the history, alerts on thresholds and reports when no data has arrived for three intervals. The agent is available for Linux, Windows and macOS. It delivers the full scope on Linux.

Disks per mount point of the server web-01: four mount points with size, used space and usage in percent, below the usage history per mount point.
Screenshot from the portal with sample data (German interface).

Server monitoring in the portal

  • Offline alert after three intervals plus a 90-second grace period, with an all-clear.
  • Threshold alerts on CPU, RAM, swap and disk, also per drive.
  • The agent opens no port and only sends outbound over port 443.

What PeakNode measures and reports per server

In PeakNode, one server monitor stands for one server. If it has GPUs, their values are added.

CPU, RAM, swap and system load

The Linux agent measures CPU utilization over a measurement window rather than from a snapshot. For RAM, used means total minus available; the page cache (the cache for files) is shown separately. Added to this are swap and the system load over 1, 5 and 15 minutes.

Every drive individually

The agent reports every mounted file system with size, usage and free space, including an NFS share. “/” and up to 32 of the fullest are stored. If a network drive does not respond, the report still goes out, with the reason instead of an empty list.

Network throughput without double counting

Download and upload as a rate, plus the totals since the last system start. Only physical interfaces are counted; in a container, all except loopback and the bridges of the container runtime.

Alert when a server stops sending data

If no data arrives for three check intervals and a 90-second grace period, the server counts as offline and the alert goes to the monitor’s channels. When it reports again, the all-clear follows.

Threshold alerts that tolerate short spikes

You set alerts on CPU, RAM, swap and disk, also for a single drive, each as “higher than” or “lower than”. Each rule can optionally trigger only once it is met in 1 to 10 consecutive checks. There are no threshold alerts on network and load.

Correct values in containers too

In Docker and Kubernetes the agent reports the assigned memory limit and CPU utilization relative to the quota, not the host’s values (cgroup v1 and v2). Without cron in the container, it runs as a background loop that lasts until the next restart.

Disks per mount point

Example

Mount pointSizeUsedUsage
/438 GiB96 GiB22%
/data7.0 TiB6.1 TiB87%
/mnt/datasets (NFS)20.0 TiB11.4 TiB57%
/var/lib/docker1.8 TiB0.7 TiB39%

Columns as in the portal. A mount point is the path under which a drive is mounted.

All readings and their sources

Sources of the Linux agent. If a system does not provide a value, the portal shows “-” instead of a zero.

WhatSourceWhen reported
CPU utilization/proc/stat, in containers the cgroup countersThreshold alert configurable
RAM used, available, page cache/proc/meminfo, in containers the cgroup limitThreshold alert on RAM utilization
Swap/proc/meminfoThreshold alert configurable
Disk (root file system)df -Pk /Threshold alert configurable
Disk per mount point: size, used, free, file systemdf -P -k, in the background with a timeoutThreshold alert per mount point or for any, then the fullest counts
Network: download, upload, totals since system start/proc/net/dev, physical interfaces onlyDisplay and history, no threshold alert
System load (1, 5, 15 minutes)/proc/loadavgDisplay and history, no threshold alert
System: operating system, kernel, CPU model, cores, architecture, uptime/etc/os-release, uname, /proc/cpuinfo, lscpu, /proc/uptimeDisplay
CPU temperature and inlet temperatureSystem sensors (hwmon) and BMCThreshold alert configurable
Memory per DIMM: corrected and uncorrectable errorsKernel EDACItem in NodeCare
Drives: health, wear, spare capacity and temperatureSMARTItem in NodeCare
Fans and power suppliesBMC, on workstations the system sensorsItem in NodeCare on failure
Server no longer reportingPortal: time of the last reportOffline alert after three intervals plus 90 seconds, all-clear on return

The hardware values (temperatures, memory, drive health, fans, power supplies) are provided by the Linux agent version 11 or later, where the system makes them available. If a source is missing, the portal shows “-”. Not part of server monitoring: a process list of the server and application logs. Whether a service responds from outside is checked by uptime monitoring. Which processes are behind a CPU spike is shown only by NodeGuard.

What the agent provides per operating system

The Linux agent is the complete one. The agents for Windows and macOS cover the system.

ReadingLinuxWindowsmacOS
CPU utilization/proc/stat over the measurement window (with NVIDIA GPUs up to 285 seconds, otherwise 3 seconds)Processor idle counters over 3 secondstop, two measurements
RAMtotal minus available, page cache separatetotal minus freeactive, wired and compressed pages
Swapyesnono
System loadyesnoyes
Drivesevery file system from dfC: as the main value, plus local drives with letters, no network drivesVolumes from df, without helper volumes and read-only volumes
Networkphysical interfacesphysical adaptersen interfaces
Basic GPU values via nvidia-smiyes, as a mean over a measurement windowyes, as a snapshotno
ECC, PCIe, throttle reason, Xid, processes per GPUyesnono
Data for NodeGuardyes, if NodeGuard is switched on for the servernono
StartCron, without cron as a background loopScheduled task under SYSTEMCron

Requirements according to the portal: on Windows, version 10 or 11 and PowerShell with administrator rights; on macOS, Big Sur or later and English as the system language. The Linux agent needs curl and bash.

Set up in three steps

No service and no inbound port. The agent sends outbound over port 443, the portal receives.

  1. Create a server monitor

    Enter a name and address or identifier and choose the check interval: 1, 5, 10, 15 or 30 minutes, depending on the plan. If you change the interval later, you need to reinstall the agent.

  2. Run one line on the server

    The portal shows one line to copy for each operating system. On Linux, the agent is then located under /opt when installed as root, otherwise in the home directory, and is entered in the crontab.

  3. Set alerts and channels

    Switch on the offline alert, set thresholds and choose where notifications go: email, Microsoft Teams, Slack, Telegram, Discord or webhook.

Frequently asked
questions about Server monitoring

Continue to:

What does PeakNode server monitoring measure?

Per server: CPU utilization, RAM, swap, system load, the usage of every drive and network throughput. If the server has NVIDIA GPUs, their values are added, see GPU monitoring.

When does PeakNode report a server as offline?

After three missed check intervals plus a 90-second grace period. With a one-minute interval, that is 270 seconds after the last report.

Can PeakNode monitor every drive individually, including an NFS share?

Yes. The agent reports every mounted file system, so /data or an NFS share as well. It only leaves out tmpfs, overlay, squashfs and mounted CD images. An alert applies to a specific mount point or to any, in which case the fullest counts.

Does the PeakNode agent also run on Windows and macOS?

Yes, with a smaller scope. The Windows agent reports CPU, RAM, the local drives, network and the basic values of NVIDIA GPUs, but neither swap nor system load nor hardware health. The macOS agent reports CPU, RAM, load, volumes and network and no GPU values. NodeGuard requires the Linux agent.

How do I avoid false alerts from short load spikes?

Every alert rule has the setting “Trigger after X checks”, from 1 to 10. The alert only fires once the rule is met that many times in a row. For planned work there are maintenance windows, during which offline and threshold alerts pause.

What does server monitoring with PeakNode cost?

The Proof of Concept plan is free and includes 1 server monitor for testing. Premium costs €149 per month and includes 10, Enterprise €490 per month and 50. One server monitor is one server, regardless of how many GPUs it has. NodeCare is included in every plan, NodeGuard from Premium upwards.

What is server monitoring?

Server monitoring is the continuous monitoring of servers: CPU and RAM utilization, drive usage, network and availability, with alerts on thresholds and outages. In PeakNode an agent on every server reports its values to the portal, which stores the history and notifies the channels. Whether a service responds from outside is checked by uptime monitoring.

How do I monitor a Linux server?

With a server monitor in PeakNode and one line on the server: it downloads the agent, a shell script, and adds it to the crontab. It only needs bash, curl and awk, only sends outbound and opens no port. For the kernel log and the processes of all users, the cron job should run as root. What the agent sends is listed under security.

How do I monitor network traffic on a Linux server?

The PeakNode agent measures the network traffic of a Linux server as download and upload rates across the physical interfaces, plus the totals since boot. PeakNode does not monitor switches, routers or the traffic of other devices. Whether a device is reachable can be checked with a ping or port monitor, as long as it is publicly reachable.

Your first server in the portal in five minutes

Create an account, create a server monitor, run one line on the server. The Proof of Concept plan is free.