Changeset 4763 for trunk/Ohana/src/opihi/scripts/psched.pro

Timestamp:

Aug 14, 2005, 10:33:53 AM (21 years ago)

Author:

eugene

Message:

cleanup up communications for speed

File:

• trunk/Ohana/src/opihi/scripts/psched.pro (modified) (3 diffs)

trunk/Ohana/src/opihi/scripts/psched.pro

@@ -1 +1 @@
 
 controller exit true
-controller host add kiawe
-controller host add alala
+# controller host add kiawe
+$Ntest = 0
+# controller host add alala
 # verbose on
+pulse 1000
+controller pulse 1000
+
+macro load.machines
+  if ($0 != 2)
+    echo "load.machines (nmach)"
+    break
+  end
+
+  for i 0 $1
+    $n = $i + 1
+    sprintf host "po%02d" $n
+    controller host add $host
+  end
+end
 
 task test
   command partest
-  periods -poll 0.1
-  periods -exec 0.1
+  # polling period is no longer valid: we check for completed controller tasks
+  # correction: still valid for local tasks
+  periods -poll 0.20
+  periods -exec 0.001
   periods -timeout 10.0
-  nmax 5
+  nmax 1024
   host anyhost
 
@@ -18 +36 @@
     queuedelete stdout
     queuedelete stderr
-    memory leaks
+    date date
+    queuepush done "$date"
+    $Ntest ++
+#   memory leaks
 #   queuesize stdout -var Nstdout
 #    for i 0 $Nstdout
@@ -42 +63 @@
   end
 end
+
+# pulse == 100ms
+# poll/exit = 0.2  : 29 sec / 100 jobs
+# poll/exit = 0.1  : 20 sec / 100 jobs
+# poll/exit = 0.05 : 17 sec / 100 jobs
+# poll/exit = 0.01 : 18 sec / 100 jobs
+
+# pulse == 10ms
+# poll/exit = 0.2  : 20 sec / 100 jobs
+# poll/exit = 0.10 : 12 sec / 100 jobs
+# poll/exit = 0.05 : 12 sec / 100 jobs
+# poll/exit = 0.01 :  9 sec / 100 jobs
+
+# we are limited here by how quickly we can send data to the 
+# controller.  this is limited by the occasional 'CheckSystem'
+# loops, with ~40ms minimum.
+
+# seems to be faster on po01 from kiawe (less interference?)
+
+# pulse == 1ms, controller pulse == 1ms
+# poll/exit = 0.01 :  3 sec / 100 jobs
+# 2 mach, 3 sec
+# 4 mach, 3 sec
+# 8 mach, 3 sec
+
+# 16 machines, 500 jobs, 13 sec: 26ms / job
+# 32 machines, 1024 jobs, 26 sec: 26ms / job
+# job harvesting rate is still the limitation.  Each job harvest requires:
+#  - jobstack exit
+#  - stdout
+#  - stderr
+#  - delete
+#  - jobstack crash
+