Re: [slurm-dev] Information about suspended nodes lost after restart

[jette]

Peter,

This small attached patch will do what you request. This is only  
applicable with a configuration of FastSchedule=0 and select/linear  
and no gang scheduling. Both the select/cons_res plugin and the gang  
scheduling module rely upon bitmaps to keep track of every CPU in the  
system, so that information can not change (i.e. the CPU count must be  
defined in slurm.conf for every node when slurmctld starts and not  
change).

Quoting Peter Kruse <p...@q-leap.de>:

Hello,

I've just successfully configured power save mode in one of our clusters,
when I noticed that the information about suspended nodes is lost
when I restart slurmctl.  For example "CoresPerSocket" is set to 1
when in fact it is 4 for some nodes.  This makes it impossible to
start a job requesting the 4 Cores.
I know that I can manually enter the number of cores, Real Memory
and alike to slurm.conf, but I hoped that there is a way to automate
this process.  I also thought that this information is stored in
the database?  Is there any way to automatically make the
information about the nodes persistent?

Thanks,

  Peter



diff --git a/src/slurmctld/node_mgr.c b/src/slurmctld/node_mgr.c
index edc8dec..c968cd7 100644
--- a/src/slurmctld/node_mgr.c
+++ b/src/slurmctld/node_mgr.c
@@ -430,6 +430,14 @@ extern int load_all_node_state ( bool state_only )
 						hostset_insert(hs, node_name);
 					else
 						hs = hostset_create(node_name);
+					/* Recover hardware state for powered
+					 * down nodes */
+					node_ptr->cpus          = cpus;
+					node_ptr->sockets       = sockets;
+					node_ptr->cores         = cores;
+					node_ptr->threads       = threads;
+					node_ptr->real_memory   = real_memory;
+					node_ptr->tmp_disk      = tmp_disk;
 				}
 				if (node_state & NODE_STATE_POWER_UP) {
 					if (power_save_mode) {