Australian cloud for business software — built and run in Sydney since 2002

NetLeverage
NetLeverage

ESXi "RAM disk 'tmp' is full.": find what filled it and clear it safely

Technical article
VMware

About this article: This article is reviewed against the cited sources. This is source research, not lab validation. Check the stated versions and your own requirements before making changes.

Broadcom's knowledge base describes /tmp on an ESXi host as a memory-backed ramdisk, /var as "typically mounted as a ramdisk", and a ramdisk named 'root'. When one fills, the host cannot write to it, and you may see events such as "RAM disk 'tmp' is full." or "The ramdisk 'root' is full. As a result, the file [PATH] could not be written." The fix is to measure which ramdisk is full with vdf -h, find the largest files, remove only files you can identify as obsolete, and then deal with whatever created them.

What the symptom means

Broadcom KB 429012, "ESXi Ramdisk /tmp is 100% full" (Updated On: 08-11-2026), says /tmp "is a memory-backed ramdisk with limited capacity". Its listed symptoms are that vMotion fails to initiate, HA agents stop responding, you cannot log in to the VMware Host Client, the image profile shows 'Unknown-no profile defined', and vmkwarning.log contains a line of this shape: "WARNING: VisorFSRam: 203: Cannot extend visorfs file ... because its ramdisk (tmp) is full."

KB 422097 (Updated On: 12-12-2025, ESXi 7.0 and 8.0) describes two events that occur together: "RAM disk 'tmp' is full." and "Error (1) saving state to /bootbank". The cause it gives is that the /tmp ramdisk is out of space when the automatic backup runs.

Ramdisks can also run out of file entries rather than bytes. KB 342828 (Updated On: 09-27-2024) shows the event "The file table of the ramdisk 'tmp' is full. As a result, the file /tmp/testdir/file8202 could not be created by the application 'scp'." and lists esxcli system visorfs ramdisk list to show the ramdisks' mount points.

The ramdisks' sizes differ by ESXi version and installation. Broadcom's KBs do not give a single default size for /tmp, so always measure on your host.

Likely causes

Cause What you see How to confirm
Large or numerous files left in /tmp vdf shows tmp at or near 100% ls -lSh /tmp | head -n 20 lists the largest entries (KB 429012)
Leftover installation files, or temporary files from add-on scripts Stale FDM VIB installation files, or vim-cmd*.txt files from NSX-T or auto-backup.sh scripts, in /tmp The listing shows them (KB 429012)
Logs redirected into the ramdisk (ESXi 8.0 and newer) "The ramdisk 'root' is full" within days; the root ramdisk is smaller when logs were meant to live on persistent storage cd /scratch/log and note whether the resulting path is persistent storage (KB 394550)
Large logs in non-redirected directories such as /var vdf shows /var or another ramdisk full vdu -ah, du -sh * and find . -type f -size +10M in the suspect directory (KB 377985)
Too many files rather than too many bytes "The file table of the ramdisk 'tmp' is full" esxcli system visorfs ramdisk list (KB 342828)

Prove the cause before changing anything

All of these commands only read. Connect to the host with SSH or the console.

1. Measure the ramdisks.

vdf -h

KB 394550 and KB 377985 use vdf to show ramdisk usage. To see only tmp, KB 429012 pipes it: vdf -h | grep tmp. KB 318795 (an ESXi 6.x and 7.x article) shows a sample line tmp 192M 192M 0B 100% --; treat that as one example of the format, not as a size to expect.

2. List the largest files in /tmp.

ls -lSh /tmp | head -n 20

KB 429012 uses this command to identify the largest files or directories. KB 318795 uses du -a for the same purpose on a directory path: du -a PATH | sort -n -r | head -n 20.

3. For /var or another ramdisk, follow KB 377985: change into the directory, use ls -lh to see which entries are symbolic links (redirected to persistent storage) and which are not, then du -sh * to find the largest directories, then find . -type f -size +10M to list files over 10 MB. Only directories that are not redirected use ramdisk space.

4. Check where logs are written. On ESXi 8.0 and newer, KB 394550 says to run cd /scratch/log and note whether the result is a persistent storage device. If it is not, logging is going to the ramdisk.

5. List the ramdisks' mount points if the event mentions "file table": esxcli system visorfs ramdisk list (KB 342828 says the command shows the mount points and might help locate extraneous files).

A good result is a clear single cause: one large file, one directory of stale files, or logging pointed at the ramdisk. A bad result is a full ramdisk with no file you can identify. In that case, stop and contact Broadcom Support (KB 429012's last line) rather than deleting by guess.

Fix

KB 318795 warns: "Make sure these files are not being used before deleting."

Stale files in /tmp

KB 429012 gives three cases: stale FDM VIB installation files, vim-cmd*.txt files from the NSX-T exporter, and other files in /tmp.

Risk: a deleted file in a ramdisk cannot be recovered. Save first: run the ls -lSh listing again, record the exact path and size of each file you will remove, and confirm which process or person created it. Delete one named path at a time (rm followed by the full path from the listing), and never use a wildcard. If you cannot identify a file, leave it and escalate. Back out: there is no back out for a deleted file, which is why identification comes first.

Stale FDM VIB installation files found in /tmp can be removed. If vim-cmd*.txt files from the NSX-T exporter are present, KB 429012 says to restart the service with /etc/init.d/nsx-exporter restart and then remove the files. For other files, remove confirmed obsolete log files or misplaced files with rm.

Disruptive step: restarting the NSX-T exporter service stops its data collection until it returns. Save first: run ls -lSh /tmp | head -n 20 and keep the listing. KB 429012 gives no status command for the service, so the restart has no further save-first step; the decision to restart belongs to the owner of the NSX deployment, and the NSX documentation for your version covers the service. Back out: a service restart cannot be undone; if the service does not return, follow the NSX documentation for your version or contact Broadcom Support.

Logs sent to the ramdisk

For the 'root' is full case on ESXi 8.0 and newer, KB 394550 says to redirect logs to a persistent volume following Broadcom's "System logs are stored on non-persistent storage" page, and then reboot the host to clear the old log files out of the ramdisk. The same KB says the logging change takes effect immediately, while existing log files stay in the ramdisk until a reboot. Broadcom KB 317690, which covers that alarm, says to open the host in the vSphere Client, then Configure, System, Advanced System Settings, and ensure Syslog.global.logDir points to a persistent location. If that field is empty or points to a scratch partition, ScratchConfig.CurrentScratchLocation must show a location on persistent storage. The format for a datastore location is [Datastorename]/foldername. KB 317690 also says a vSAN datastore must not be used for this, because redirecting system logs to a vSAN object causes a host lock up.

Disruptive step: a host reboot interrupts every VM on the host. Put the host in maintenance mode and let vMotion move the VMs away first; if the full /tmp is what is stopping vMotion, free space first. Save first: in Advanced System Settings, write down the current values of Syslog.global.logDir and ScratchConfig.CurrentScratchLocation exactly as shown (a screenshot also works). Back out: open the same Advanced System Settings screen, set Syslog.global.logDir back to the value you wrote down, and save.

A reboot as a stopgap

KB 429012 says: "A host reboot will clear the /tmp ramdisk temporarily, but identifying the root cause is necessary to prevent recurrence." KB 377985 says that if you cannot access the host for normal troubleshooting, a reboot clears the ramdisk.

Disruptive step: a reboot interrupts every VM on the host, and anything held only in the ramdisk is lost. Save first: note the VMs on the host and their power state, and copy the ls -lSh /tmp | head -n 20 listing off the host (paste it into your ticket or notes) before the reboot, so the cause can still be traced. Do not write it to a file on the host, because that file would sit in the full ramdisk and be lost with the reboot. Move VMs off first if the host allows it. Back out: none; power on any VM that does not restart.

Management agents

KB 318795 ends its steps, after the unwanted files are deleted, with restarting management services. Do this only if the host still looks disconnected or unresponsive once space is freed. KB 320280 says "Do not use the services.sh restart command or Restart Management agents in ESXi Using Direct Console User Interface (DCUI)" on hosts that run vSAN, LACP, NSX or shared graphics. Its line "Only restart single services." follows the shared-graphics check, and it gives /etc/init.d/hostd restart and /etc/init.d/vpxa restart as the single-service restarts. The KB does not say whether that line applies to the vSAN, LACP and NSX cases as well.

Disruptive step: restarting hostd or vpxa can make the host appear disconnected in vCenter for a short time and can affect running tasks. KB 320280 also says a restart can lead to system instability requiring a host reboot. Save first: note the host's current connection state in the vSphere Client and the list of tasks running on it. Restart one service, wait one to two minutes, and check the host status in the vSphere Client before restarting another. Back out: a restart cannot be undone; if the host does not return to "Connected", read /var/log/hostd.log, as KB 320280 says, before any further restart.

Example (illustrative): On esx01, a vMotion off the host fails and the Host Client will not log in. vdf -h | grep tmp shows tmp at 100%. ls -lSh /tmp | head -n 20 lists a 150 MB file called collect-output-20261101.txt, left by a support script a colleague ran three weeks ago. The colleague confirms it can go. The administrator records the path, removes that one file, runs vdf -h | grep tmp again and sees usage fall, then retries the vMotion.

Confirm the fix

  1. vdf -h shows the ramdisk well below 100%, and vdf -h | grep tmp is the line KB 429012 uses to verify it.
  2. The failing action works: vMotion starts, the HA agent reports normally, the Host Client login succeeds.
  3. No new "ramdisk" lines appear in vmkwarning.log.
  4. If you changed log destination, cd /scratch/log shows a persistent path.

What to watch afterwards

  • Check vdf -h periodically after changes. KB 394550 advises monitoring ramdisk usage this way.
  • If the same file name returns, find the script, agent or add-on that creates it.
  • Compare the host with others in the cluster: a host that fills while its peers do not has a local cause.

Risks and limits

  • These commands have not been run or lab-tested.
  • The Broadcom KBs cover different ESXi versions; their dates and versions are in the Sources list. Check the one that matches your build.
  • Ramdisk names and sizes vary by installation, and no default size is given here.
  • Files in a ramdisk that belong to ESXi itself must not be removed. If a file is not clearly a leftover, ask Broadcom Support.

Sources