#
# Mark Huang <mlhuang@cs.princeton.edu>
# Andy Bavier <acb@cs.princeton.edu>
+# Faiyaz Ahmed <faiyaza@cs.princeton.edu>
# Copyright (C) 2004-2006 The Trustees of Princeton University
#
-# $Id: swapmon.py,v 1.2 2006/04/28 19:32:18 mlhuang Exp $
+# $Id: swapmon.py,v 1.9 2006/07/19 19:40:55 faiyaza Exp $
#
import syslog
change_thresh = 5
# Swap utilization at which the largest consumer of physical memory is reset
-reset_thresh = 85
+reset_thresh = 80
# Swap utilization at which the machine is rebooted
reboot_thresh = 95
-# Minimum physical memory utilization to be considered the largest consumer
-min_thresh = 10
+# Time to wait before checking slice again after reset
+reset_timeout = 15
+
+# Number of strikes before killing (strike, strike, kill)
+kill_thresh = 2
+
+# Time to wait before removing slice from kill queue (probation)
+kill_timeout = 120
+
+# Don't email the same message more than once in the same emailtimeout interval
+email_timeout = 1800
+
+# Physical size threshold to be considered a consumer. Rationale is if there are no procs
+# with a size at least as large as this, then there is a slow leaker; better to just reboot.
+rss_min = 150 * 1024
# System slices that should not be reset (regexps)
system_slices = ['root', PLC_SLICE_PREFIX + '_']
%(date)s %(hostname)s alarm %(slice)s
""".lstrip()
+# Message sent after a slice has been killed
+kill_subject = "pl_mom killed slice %(slice)s on %(hostname)s"
+kill_body = \
+"""
+Sometime before %(date)s, swap space was
+nearly exhausted on %(hostname)s.
+
+Slice %(slice)s was killed since it was the largest consumer of
+physical memory at %(rss)s (%(percent)4.1f%%) after repeated restarts.
+
+Please reply to this message explaining the nature of your experiment,
+and what you are doing to address the problem.
+
+%(slice)s processes prior to reset:
+
+%(table)s
+
+%(date)s %(hostname)s reset %(slice)s
+""".lstrip()
+
+
+
+class Reset:
+ """
+ Keeps track of state information for resets and kills
+
+ resettimeleft - timeout before checking for next reset
+ resetcount - number of strikes
+ killtimeleft - time out before removing from kill queue
+ {kill,reset}mail - Time of last email
+ kill - State of kill. If slice is already being killed, wait before retry.
+ """
+
+ def __init__(self,name):
+ self.name = name
+ self.resettimeleft = reset_timeout
+ self.resetcount = 0
+ self.resetmail = 0
+ self.killtimeleft = kill_timeout
+ self.killmail = 0
+
+ def __repr__(self):
+ return self.name
+
+ def update(self):
+ # Count down for next check of reset slice.
+ if self.resettimeleft > 0:
+ self.resettimeleft -= 1
+ if debug and verbose:
+ print "%s has %s seconds in probation" \
+ %(self.name, self.killtimeleft)
+ if self.killtimeleft > 0:
+ # Count down kill probation timer (killtimeleft)
+ self.killtimeleft -= 1
+ if self.killtimeleft == 1:
+ print "%s is out of probation" % self.name
+ else:
+ # Once out of probation period (killtimeleft), remove strikes
+ self.resetcount = 0
+
+
+ # Check to see if a slice needs to be killed. If it has been killed more
+ # than kill_thresh in the probation period (kill_timeout) send an email, kill the slice.
+ def checkkill(self,params):
+ if self.killtimeleft > 0 and self.resetcount >= kill_thresh:
+ if debug:
+ print kill_subject % params
+ print kill_body % params
+ try:
+ pid = os.fork()
+ if pid == 0:
+ print "Slice %s is being killed." % self.name
+ vserver = VServer(self.name)
+ vserver.stop()
+ os._exit(0)
+ else:
+ os.waitpid(pid,0)
+ except Exception, err:
+ print "Warning: Exception received while killing slice %s: %s" \
+ % self.name, err
+ if (time.time() - self.killmail) > email_timeout:
+ slicemail(self.name, kill_subject % params, kill_body % params)
+ print "Sending KILL email for slice %s" % self.name
+ self.killmail = time.time()
+ return True
+ return False
+
+ # Reset slice after checking to see if slice is out of timeout.
+ # Increment resetcount, check to see if larger than kill_thresh.
+ def reset(self, params):
+ # If its the first reset (came back after kill)
+ # or if its been reset before
+ # and we are out of the reset timeout.
+ if self.resetcount == 0 or self.resettimeleft == 0:
+ # Do we need to kill this slice? Check history first.
+ if self.checkkill(params):
+ return
+ # Update counters
+ self.resetcount += 1
+ self.killtimeleft = kill_timeout
+ self.resettimeleft = reset_timeout
+ print "%s has %s seconds to die and has been reset %s times" \
+ %(self.name, self.resettimeleft, self.resetcount)
+ if debug:
+ print reset_subject % params
+ print reset_body % params
+ try:
+ pid = os.fork()
+ if pid == 0:
+ print "Resetting slice " + self.name
+ vserver = VServer(self.name)
+ vserver.stop()
+ vserver.start(wait = False)
+ os._exit(0)
+ else:
+ os.waitpid(pid,0)
+ except Exception, err:
+ print "Warning: Exception received while resetting slice %s:" \
+ % self.name, err
+ if (time.time() - self.resetmail) > email_timeout:
+ slicemail(self.name, reset_subject % params, reset_body % params)
+ print "Sending Reset email for slice %s" % self.name
+ self.resetmail = time.time()
+
+
def usage():
print """
Usage: %s [OPTIONS]...
except ValueError:
pass
+ # vps sometimes prints ERR instead of a context ID if it
+ # cannot identify the context of an orphaned (usually dying)
+ # process. Skip these processes.
+ if type(proc['xid']) != int:
+ continue
+
# Assign (pl_)sshd processes to slice instead of root
m = re.search(r"sshd: ([a-zA-Z_]+)", proc['cmd'])
if m is not None:
def main():
# Defaults
global debug, verbose, datafile
- global period, change_thresh, reset_thresh, reboot_thresh, min_thresh, system_slices
+ global period, change_thresh, reset_thresh, reboot_thresh, rss_min, system_slices
# All slices
names = []
reset_thresh = int(optval)
elif opt == "--reboot-thresh":
reboot_thresh = int(optval)
- elif opt == "--min-thresh":
- min_thresh = int(optval)
elif opt == "--system-slice":
system_slices.append(optval)
elif opt == "--status":
(version, slices) = pickle.load(f)
f.close()
# Check version of data file
- if version != "$Id: swapmon.py,v 1.2 2006/04/28 19:32:18 mlhuang Exp $":
+ if version != "$Id: swapmon.py,v 1.9 2006/07/19 19:40:55 faiyaza Exp $":
print "Not using old version '%s' data file %s" % (version, datafile)
raise Exception
# Delete data file
os.unlink(datafile)
except Exception:
- version = "$Id: swapmon.py,v 1.2 2006/04/28 19:32:18 mlhuang Exp $"
+ version = "$Id: swapmon.py,v 1.9 2006/07/19 19:40:55 faiyaza Exp $"
slices = {}
# Query process table every 30 seconds, or when a large change in
# System slices that we have warned but could not reset
warned = []
+ # Slices that were reset
+ resetlist = {}
+
while True:
used = swap_used()
+
+ for resetslice in resetlist.keys():
+ resetlist[resetslice].update()
+
if last_used is None:
last_used = used
- if verbose:
+
+ if verbose:
print "%d%% swap consumed" % used
if used >= reboot_thresh:
bwlimit.run("/bin/sync; /sbin/reboot -f")
elif used >= reset_thresh:
+ if debug:
+ print "Memory used = %s" %(used)
# Try and find a hog
slicelist = slices.values()
slicelist.sort(lambda a, b: b['rss'] - a['rss'])
for slice in slicelist:
percent = 100. * slice['rss'] / total_rss
- if percent < min_thresh:
- continue
+ if slice['rss'] < rss_min:
+ continue
+
print "%d%% swap consumed, slice %s is using %s (%d%%) of memory" % \
(used,
slice['name'],
# Match slice name against system slice patterns
is_system_slice = filter(None, [re.match(pattern, slice['name']) for pattern in system_slices])
- if is_system_slice:
- # Do not reset system slices, just warn once
- if slice['name'] not in warned:
- warned.append(slice['name'])
- if debug:
- print alarm_subject % params
- print alarm_body % params
- else:
- print "Warning slice " + slice['name']
- slicemail(slice['name'], alarm_subject % params, alarm_body % params)
+ if is_system_slice:
+ if slice['name'] not in warned:
+ warned.append(slice['name'])
+ if debug:
+ print alarm_subject % params
+ print alarm_body % params
+ else:
+ print "Warning slice " + slice['name']
+ slicemail(slice['name'], alarm_subject % params,
+ alarm_body % params)
else:
- # Otherwise, reset
- if debug:
- print reset_subject % params
- print reset_body % params
- else:
- try:
- pid = os.fork()
- if pid == 0:
- print "Resetting slice " + slice['name']
- vserver = VServer(slice['name'])
- vserver.stop()
- vserver.start(wait = False)
- os._exit(0)
- else:
- os.waitpid(pid, 0)
- except Exception, err:
- print "Warning: Exception received while resetting slice %s:" % slice['name'], err
- slicemail(slice['name'], reset_subject % params, reset_body % params)
- break
-
- elif timer <= 0 or used >= (last_used + change_thresh):
+ # Reset slice
+ if not resetlist.has_key(slice['name']):
+ resetlist[slice['name']] = Reset(slice['name'])
+ resetlist[slice['name']].reset(params)
+ slices = slicestat(names)
+
+ if timer <= 0 or used >= (last_used + change_thresh):
if used >= (last_used + change_thresh):
print "%d%% swap consumed, %d%% in last %d seconds" % \
(used, used - last_used, period - timer)