m/fox/configuration.nix

@@ -93,4 +93,20 @@
     wantedBy = [ "multi-user.target" ];
     serviceConfig.ExecStart = script;
   };
+
+  # Only allow SSH connections from users who have a SLURM allocation
+  # See: https://slurm.schedmd.com/pam_slurm_adopt.html
+  security.pam.services.sshd.rules.account.slurm = {
+    control = "required";
+    enable = true;
+    modulePath = "${pkgs.slurm}/lib/security/pam_slurm_adopt.so";
+    args = [ "log_level=debug5" ];
+    order = 999999; # Make it last one
+  };
+
+  # Disable systemd session (pam_systemd.so) as it will conflict with the
+  # pam_slurm_adopt.so module. What happens is that the shell is first adopted
+  # into the slurmstepd task and then into the systemd session, which is not
+  # what we want, otherwise it will linger even if all jobs are gone.
+  security.pam.services.sshd.startSession = lib.mkForce false;
 }

m/module/slurm-client.nix

@@ -1,4 +1,4 @@
-{ lib, pkgs, ... }:
+{ lib, ... }:
 
 {
   imports = [
@@ -21,20 +21,4 @@
   };
 
   services.slurm.client.enable = true;
-
-  # Only allow SSH connections from users who have a SLURM allocation
-  # See: https://slurm.schedmd.com/pam_slurm_adopt.html
-  security.pam.services.sshd.rules.account.slurm = {
-    control = "required";
-    enable = true;
-    modulePath = "${pkgs.slurm}/lib/security/pam_slurm_adopt.so";
-    args = [ "log_level=debug5" ];
-    order = 999999; # Make it last one
-  };
-
-  # Disable systemd session (pam_systemd.so) as it will conflict with the
-  # pam_slurm_adopt.so module. What happens is that the shell is first adopted
-  # into the slurmstepd task and then into the systemd session, which is not
-  # what we want, otherwise it will linger even if all jobs are gone.
-  security.pam.services.sshd.startSession = lib.mkForce false;
 }

m/module/slurm-common.nix

@@ -1,6 +1,31 @@
 { config, pkgs, ... }:
 
-{
+let
+  suspendProgram = pkgs.writeShellScript "suspend.sh" ''
+    exec 1>>/var/log/power_save.log 2>>/var/log/power_save.log
+    set -x
+    export "PATH=/run/current-system/sw/bin:$PATH"
+    echo "$(date) Suspend invoked $0 $*" >> /var/log/power_save.log
+    hosts=$(scontrol show hostnames $1)
+    for host in $hosts; do
+      echo Shutting down host: $host
+      ipmitool -I lanplus -H ''${host}-ipmi -P "" -U "" chassis power off
+    done
+  '';
+
+  resumeProgram = pkgs.writeShellScript "resume.sh" ''
+    exec 1>>/var/log/power_save.log 2>>/var/log/power_save.log
+    set -x
+    export "PATH=/run/current-system/sw/bin:$PATH"
+    echo "$(date) Suspend invoked $0 $*" >> /var/log/power_save.log
+    hosts=$(scontrol show hostnames $1)
+    for host in $hosts; do
+      echo Starting host: $host
+      ipmitool -I lanplus -H ''${host}-ipmi -P "" -U "" chassis power on
+    done
+  '';
+
+in {
   services.slurm = {
     controlMachine = "apex";
     clusterName = "jungle";
@@ -34,6 +59,16 @@
       # the resources. Use the task/cgroup plugin to enable process containment.
       TaskPlugin=task/affinity,task/cgroup
 
+      # Power off unused nodes until they are requested
+      SuspendProgram=${suspendProgram}
+      SuspendTimeout=60
+      ResumeProgram=${resumeProgram}
+      ResumeTimeout=300
+      SuspendExcNodes=fox
+
+      # Turn the nodes off after 1 hour of inactivity
+      SuspendTime=3600
+
       # Reduce port range so we can allow only this range in the firewall
       SrunPortRange=60000-61000