diff --git a/docs/source/troubleshooting/os_installation/index.rst b/docs/source/troubleshooting/os_installation/index.rst
index 6a263c5f3..ac8ed52cf 100644
--- a/docs/source/troubleshooting/os_installation/index.rst
+++ b/docs/source/troubleshooting/os_installation/index.rst
@@ -41,4 +41,5 @@ Next chapter introduces the procedures on how to troubleshoot operating system i
    log_to_mn_cn.rst
    debug_port.rst
    ssh_enable.rst
+   ubuntu_subiquity_memory.rst
 
diff --git a/docs/source/troubleshooting/os_installation/ubuntu_subiquity_memory.rst b/docs/source/troubleshooting/os_installation/ubuntu_subiquity_memory.rst
new file mode 100644
index 000000000..c31079284
--- /dev/null
+++ b/docs/source/troubleshooting/os_installation/ubuntu_subiquity_memory.rst
@@ -0,0 +1,28 @@
+Ubuntu diskful install needs memory for the live filesystem
+===========================================================
+
+An Ubuntu diskful (Subiquity) install boots the live installer over NFS and then copies it
+into RAM, because the kernel command line xCAT generates carries ``toram``. This is
+deliberate: with the NFS root still mounted at the end of the install, ``systemd-shutdown``
+blocks on I/O to it and the node never reboots into the disk it just installed.
+
+The consequence is that the compute node must have enough memory for the live filesystem on
+top of whatever the installer itself needs. The squashfs layers are roughly 1.5 GB on 24.04
+and grow with each release, so **4 GB is a practical floor and 8 GB is comfortable**.
+
+A node with too little memory fails part-way through the install, and the symptom resembles
+an unrelated problem: the node reboots back into the installer, so it looks like a PXE loop
+rather than an out-of-memory condition. Check the console before assuming a boot-flip
+failure. ::
+
+    rcons <node>
+
+To confirm, look for the OOM killer in the installer's kernel messages, or watch the node's
+memory from the hypervisor while it installs. For a libvirt-hosted node the memory is
+``vmm``/``vmmemory`` in the **vm** table. ::
+
+    lsdef <node> -i vmmemory
+    chdef <node> vmmemory=8192
+
+Diskless (netboot) nodes have a related but separate requirement: the rootimg is unpacked
+into a tmpfs, and on ppc64le the 64 KB page size inflates it considerably.
diff --git a/xCAT-server/lib/xcat/plugins/debian.pm b/xCAT-server/lib/xcat/plugins/debian.pm
index 1b039836d..a4ac3a6ef 100644
--- a/xCAT-server/lib/xcat/plugins/debian.pm
+++ b/xCAT-server/lib/xcat/plugins/debian.pm
@@ -524,7 +524,10 @@ sub copycd
     network root at shutdown. Without it a process doing I/O to it blocks uninterruptibly,
     systemd-shutdown waits forever and the node never reboots into the disk it just installed.
     casper parses only nfsroot= and takes the whole value as the path, so mount options cannot
-    be appended; toram is its supported alternative.
+    be appended; toram is its supported alternative. It does mean the node needs memory for the
+    live filesystem on top of the installer -- roughly 1.5G on 24.04 -- or it dies part-way
+    through and reboots into the installer again, which looks like a boot-flip failure. See
+    docs/source/troubleshooting/os_installation/ubuntu_subiquity_memory.rst.
 
     Arguments:
         $base       the command line built so far