Merge pull request #3126 from GoogleCloudPlatform/pr2979

Use sackd for the login nodes
GoogleCloudPlatform · Oct 12, 2024 · 52bd393 · 52bd393
2 parents dec5719 + b59f80c
commit 52bd393
Show file tree

Hide file tree

Showing 3 changed files with 14 additions and 30 deletions.
diff --git a/...ity/modules/scheduler/schedmd-slurm-gcp-v6-controller/modules/slurm_files/scripts/conf.py b/...ity/modules/scheduler/schedmd-slurm-gcp-v6-controller/modules/slurm_files/scripts/conf.py
@@ -27,7 +27,6 @@
 # This file is managed by a script. Manual modifications will be overwritten.
 """
 
-login_nodeset = "x-login"
 
 
 def dict_to_conf(conf, delim=" ") -> str:
@@ -130,24 +129,6 @@ def get(key, default):
     return dict_to_conf(conf_options, delim="\n")
 
 
-def loginlines() -> str:
-    nodeset = {
-        "NodeSet": login_nodeset,
-        "Feature": login_nodeset,
-    }
-    partition = {
-        "PartitionName": login_nodeset,
-        "Nodes": login_nodeset,
-        "State": "UP",
-        "DefMemPerCPU": 1,
-        "Hidden": "YES",
-        "RootOnly": "YES",
-    }
-    lines = [
-        dict_to_conf(nodeset),
-        dict_to_conf(partition),
-    ]
-    return "\n".join(lines)
 
 
 def nodeset_lines(nodeset, lkp: util.Lookup) -> str:
@@ -254,7 +235,7 @@ def suspend_exc_lines(lkp: util.Lookup) -> Iterable[str]:
         for p in lkp.cfg.partitions.values()
         if len(p.partition_nodeset_dyn) > 0
     ]
-    suspend_exc_parts = {"SuspendExcParts": [login_nodeset, *dyn_parts]}
+    suspend_exc_parts = {"SuspendExcParts": [*dyn_parts]}
 
     return filter(
         None,
@@ -270,7 +251,6 @@ def make_cloud_conf(lkp: util.Lookup) -> str:
     lines = [
         FILE_PREAMBLE,
         conflines(lkp),
-        loginlines(),
         *(nodeset_lines(n, lkp) for n in lkp.cfg.nodeset.values()),
         *(nodeset_dyn_lines(n) for n in lkp.cfg.nodeset_dyn.values()),
         *(nodeset_tpu_lines(n, lkp) for n in lkp.cfg.nodeset_tpu.values()),

diff --git a/...ty/modules/scheduler/schedmd-slurm-gcp-v6-controller/modules/slurm_files/scripts/setup.py b/...ty/modules/scheduler/schedmd-slurm-gcp-v6-controller/modules/slurm_files/scripts/setup.py
@@ -216,6 +216,7 @@ def setup_sudoers():
     content = """
 # Allow SlurmUser to manage the slurm daemons
 slurm ALL= NOPASSWD: /usr/bin/systemctl restart slurmd.service
+slurm ALL= NOPASSWD: /usr/bin/systemctl restart sackd.service
 slurm ALL= NOPASSWD: /usr/bin/systemctl restart slurmctld.service
 """
     sudoers_file = Path("/etc/sudoers.d/slurm")
@@ -366,27 +367,25 @@ def setup_login():
     slurmctld_host = f"{lookup().control_host}"
     if lookup().control_addr:
         slurmctld_host = f"{lookup().control_host}({lookup().control_addr})"
-    slurmd_options = [
+    sackd_options = [
         f'--conf-server="{slurmctld_host}:{lookup().control_host_port}"',
-        f'--conf="Feature={conf.login_nodeset}"',
-        "-Z",
     ]
-    sysconf = f"""SLURMD_OPTIONS='{" ".join(slurmd_options)}'"""
-    update_system_config("slurmd", sysconf)
+    sysconf = f"""SACKD_OPTIONS='{" ".join(sackd_options)}'"""
+    update_system_config("sackd", sysconf)
     install_custom_scripts()
 
     setup_network_storage()
     setup_sudoers()
     run("systemctl restart munge")
-    run("systemctl enable slurmd", timeout=30)
-    run("systemctl restart slurmd", timeout=30)
+    run("systemctl enable sackd", timeout=30)
+    run("systemctl restart sackd", timeout=30)
     run("systemctl enable --now slurmcmd.timer", timeout=30)
 
     run_custom_scripts()
 
     log.info("Check status of cluster services")
     run("systemctl status munge", timeout=30)
-    run("systemctl status slurmd", timeout=30)
+    run("systemctl status sackd", timeout=30)
 
     log.info("Done setting up login")
 

diff --git a/...odules/scheduler/schedmd-slurm-gcp-v6-controller/modules/slurm_files/scripts/slurmsync.py b/...odules/scheduler/schedmd-slurm-gcp-v6-controller/modules/slurm_files/scripts/slurmsync.py
@@ -428,11 +428,16 @@ def reconfigure_slurm():
             log.exception("failed to reconfigure slurmctld")
         util.run(f"wall '{update_msg}'", timeout=30)
         log.debug("Done.")
-    elif lookup().instance_role_safe in ["compute", "login"]:
+    elif lookup().instance_role_safe == "compute":
         log.info("Restarting slurmd to make changes take effect.")
         run("systemctl restart slurmd")
         util.run(f"wall '{update_msg}'", timeout=30)
         log.debug("Done.")
+    elif lookup().instance_role_safe == "login":
+        log.info("Restarting sackd to make changes take effect.")
+        run("systemctl restart sackd")
+        util.run(f"wall '{update_msg}'", timeout=30)
+        log.debug("Done.")
 
 
 def update_topology(lkp: util.Lookup) -> None: