Enable memory limits in SLURM

Make sure that jobs cannot allocate more memory than available so we
don't trigger the OOM killer.

Fixes: rarias/jungle#178

m/module/slurm-common.nix

@@ -5,8 +5,8 @@
     controlMachine = "apex";
     clusterName = "jungle";
     nodeName = [
-      "owl[1,2]  Sockets=2 CoresPerSocket=14 ThreadsPerCore=2 Feature=owl"
-      "fox       Sockets=8 CoresPerSocket=24 ThreadsPerCore=1"
+      "owl[1,2]  Sockets=2 CoresPerSocket=14 ThreadsPerCore=2 MemSpecLimit=4096 RealMemory=128797"
+      "fox       Sockets=8 CoresPerSocket=24 ThreadsPerCore=1 MemSpecLimit=4096 RealMemory=773659"
     ];
 
     partitionName = [
@@ -41,9 +41,9 @@
       # multiple hardware threads (or CPUs).
       SelectType=select/cons_tres
 
-      # Ignore memory constraints and only use unused cores to share a node with
-      # other jobs.
-      SelectTypeParameters=CR_Core
+      # Both cores and memory are consumable resources, so we can put a limit in
+      # memory as well.
+      SelectTypeParameters=CR_Core_Memory
 
       # Required for pam_slurm_adopt, see https://slurm.schedmd.com/pam_slurm_adopt.html
       # This sets up the "extern" step into which ssh-launched processes will be
@@ -58,7 +58,9 @@
 
     extraCgroupConfig = ''
       CgroupPlugin=cgroup/v2
-      #ConstrainCores=yes
+      ConstrainRAMSpace=yes
+      ConstrainSwapSpace=yes
+      AllowedRAMSpace=99
     '';
   };