3 # Swap monitoring daemon. Every 30 seconds, checks process memory
4 # usage. At 90% utilization, resets the slice that is consuming the
5 # most physical memory. At 95% utilization, reboots the machine to
8 # Mark Huang <mlhuang@cs.princeton.edu>
9 # Andy Bavier <acb@cs.princeton.edu>
10 # Faiyaz Ahmed <faiyaza@cs.princeton.edu>
11 # Copyright (C) 2004-2006 The Trustees of Princeton University
13 # $Id: swapmon.py,v 1.5 2006/05/09 03:23:57 mlhuang Exp $
25 # util-vserver/python/vserver.py allows us to control slices directly
27 from vserver import VServer
29 # bwlimit exports a few useful functions like run(), get_xid(), and get_slice()
38 datafile = "/var/lib/misc/swapmon.dat"
40 # Seconds between process analysis
43 # Minimum change in swap utilization over 30 seconds that will trigger
44 # early process analysis.
47 # Swap utilization at which the largest consumer of physical memory is reset
50 # Swap utilization at which the machine is rebooted
53 # Time to wait before checking slice again after reset
56 # Number of strikes before killing (strike, strike, kill)
59 # Time to wait before removing slice from kill queue (probation)
62 # Don't email the same message more than once in the same emailtimeout interval
65 # Minimum physical memory utilization to be considered the largest consumer
68 # System slices that should not be reset (regexps)
69 system_slices = ['root', PLC_SLICE_PREFIX + '_']
71 # Message sent after a critical reboot
72 rebooted_subject = "pl_mom rebooted %(hostname)s"
75 Sometime before %(date)s, swap space was
76 nearly exhausted on %(hostname)s, so pl_mom rebooted it.
78 Slices active prior to reboot are listed below. Memory usage
79 statistics are not entirely accurate due to threading.
83 %(date)s %(hostname)s reboot
86 # Message sent after a hog is reset
87 reset_subject = "pl_mom reset slice %(slice)s on %(hostname)s"
90 Sometime before %(date)s, swap space was
91 nearly exhausted on %(hostname)s.
93 Slice %(slice)s was reset since it was the largest consumer of
94 physical memory at %(rss)s (%(percent)4.1f%%).
96 Please reply to this message explaining the nature of your experiment,
97 and what you are doing to address the problem.
99 %(slice)s processes prior to reset:
103 %(date)s %(hostname)s reset %(slice)s
106 # Message sent to system slices that should not be reset
107 alarm_subject = "pl_mom alarm slice %(slice)s on %(hostname)s"
110 Sometime before %(date)s, swap space was
111 nearly exhausted on %(hostname)s.
113 System slice %(slice)s was the largest consumer of physical memory at
114 %(rss)s (%(percent)4.1f%%). It was not reset, but please verify its
117 %(slice)s processes prior to alarm:
121 %(date)s %(hostname)s alarm %(slice)s
124 # Message sent after a slice has been killed
125 kill_subject = "pl_mom killed slice %(slice)s on %(hostname)s"
128 Sometime before %(date)s, swap space was
129 nearly exhausted on %(hostname)s.
131 Slice %(slice)s was killed since it was the largest consumer of
132 physical memory at %(rss)s (%(percent)4.1f%%) after repeated restarts.
134 Please reply to this message explaining the nature of your experiment,
135 and what you are doing to address the problem.
137 %(slice)s processes prior to reset:
141 %(date)s %(hostname)s reset %(slice)s
148 Keeps track of state information for resets and kills
150 resettimeleft - timeout before checking for next reset
151 resetcount - number of strikes
152 killtimeleft - time out before removing from kill queue
153 {kill,reset}mail - Time of last email
154 kill - State of kill. If slice is already being killed, wait before retry.
157 def __init__(self,name):
159 self.resettimeleft = reset_timeout
163 self.killtimeleft = kill_timeout
170 # Count down for next check of reset slice.
171 if self.resettimeleft > 0:
172 self.resettimeleft -= 1
173 if debug and verbose:
174 print "%s has %s seconds in probation" \
175 %(self.name, self.killtimeleft)
176 if self.killtimeleft > 0:
177 # Count down kill probation timer (killtimeleft)
178 self.killtimeleft -= 1
179 if self.killtimeleft == 1:
180 print "%s is out of probation" % self.name
182 # Once out of probation period (killtimeleft), remove strikes
187 # Check to see if a slice needs to be killed. If it has rules more than kill_thresh in
188 # the probation period (kill_timeout) send an email, kill the slice.
189 def checkkill(self,params):
190 if self.killtimeleft > 0 and self.resetcount >= kill_thresh and \
194 print kill_subject % params
195 print kill_body % params
199 print "Slice %s is being killed." % self.name
200 vserver = VServer(self.name)
205 except Exception, err:
206 print "Warning: Exception received while killing slice %s: %s" \
208 if (time.time() - self.killmail) > email_timeout:
209 slicemail(self.name, kill_subject % params, kill_body % params)
210 print "Sending KILL email for slice %s" % self.name
211 self.killmail = time.time()
215 # Reset slice after checking to see if slice is out of timeout.
216 # Increment resetcount, check to see if larger than kill_thresh.
217 def reset(self, params):
218 # If its the first reset or if its been reset before
219 # and we are out of the reset timeout.
220 if self.resetcount == 0 or self.resettimeleft == 0:
221 # Do we need to kill this slice? Check history first.
222 if self.checkkill(params):
226 self.killtimeleft = kill_timeout
227 self.resettimeleft = reset_timeout
228 print "%s has %s seconds to die and has been reset %s times" \
229 %(self.name, self.resettimeleft, self.resetcount)
231 print reset_subject % params
232 print reset_body % params
236 print "Resetting slice " + self.name
237 vserver = VServer(self.name)
239 vserver.start(wait = False)
243 except Exception, err:
244 print "Warning: Exception received while resetting slice %s:" \
246 if (time.time() - self.resetmail) > email_timeout:
247 slicemail(self.name, reset_subject % params, reset_body % params)
248 print "Sending Reset email for slice %s" % self.name
249 self.resetmail = time.time()
254 Usage: %s [OPTIONS]...
257 -d, --debug Enable debugging (default: %s)
258 -v, --verbose Increase verbosity level (default: %d)
259 -f, --file=FILE Data file (default: %s)
260 -s, --slice=SLICE Constrain monitoring to these slices (default: all)
261 -p, --period=SECONDS Seconds between normal process analysis (default: %s)
262 --reset-thresh=PERCENT Swap utilization at which slice reset is attempted
263 --reboot-thresh=PERCENT Swap utilization at which the machine is rebooted
264 --min-thresh=PERCENT Minimum physical memory utilization to be considered a hog
265 --system-slice=SLICE System slice that should not be reset
266 --status Print memory usage statistics and exit
267 -h, --help This message
268 """.lstrip() % (sys.argv[0], debug, verbose, datafile, format_period(period))
270 def slicestat(names = None):
272 Get status of specified slices (if names is None or empty, all
273 slices). vsize and rss are in KiB. Returns
275 {xid: {'xid': slice_id,
277 'procs': [{'pid': pid, 'xid': slice_id, 'user', username, 'cmd': command,
278 'vsize': virtual_kib, 'rss': physical_kib,
279 'pcpu': cpu_percent, 'pmem': mem_percent}]
280 'vsize': total_virtual_kib,
281 'rss': total_physical_kib}}
284 # Mandatory fields. xid is a virtual field inserted by vps. Make
285 # sure cmd is last so that it does not get truncated
287 fields = ['pid', 'xid', 'user', 'vsize', 'rss', 'pcpu', 'pmem', 'cmd']
289 # vps inserts xid after pid in the output, but ps doesn't know
290 # what the field means.
291 ps_fields = list(fields)
292 ps_fields.remove('xid')
296 # Eat the header line. vps depends on the header to figure out
297 # which column is the PID column, so we can't just tell ps not to
299 for line in bwlimit.run("/usr/sbin/vps -e -o " + ",".join(ps_fields))[1:]:
303 # Replace "0 MAIN" and "1 ALL_PROC" (the special monikers that
304 # vps uses to denote the root context and the "all contexts"
305 # context) with "0" so that we can just split() on whitespace.
306 line = line.replace("0 MAIN", "0").replace("1 ALL_PROC", "0")
308 # Represent process as a dict of fields
309 values = line.split(None, len(fields) - 1)
310 if len(values) != len(fields):
312 proc = dict(zip(fields, values))
314 # Convert ints and floats
317 proc[field] = int(proc[field])
320 proc[field] = float(proc[field])
324 # vps sometimes prints ERR instead of a context ID if it
325 # cannot identify the context of an orphaned (usually dying)
326 # process. Skip these processes.
327 if type(proc['xid']) != int:
330 # Assign (pl_)sshd processes to slice instead of root
331 m = re.search(r"sshd: ([a-zA-Z_]+)", proc['cmd'])
333 xid = bwlimit.get_xid(m.group(1))
337 name = bwlimit.get_slice(proc['xid'])
339 # Orphaned (not associated with a slice) class
340 name = "%d?" % proc['xid']
342 # Monitor only the specified slices
343 if names and name not in names:
346 # Additional overhead calculations from slicestat
348 # Include 12 KiB of process overhead =
349 # 4 KiB top-level page table +
350 # 4 KiB kernel structure +
351 # 4 KiB basic page table
354 # Include additional page table overhead
355 if proc['vsize'] > 4096:
356 proc['rss'] += 4 * ((proc['vsize'] - 1) / 4096)
358 if slices.has_key(proc['xid']):
359 slice = slices[proc['xid']]
361 slice = {'xid': proc['xid'], 'name': name, 'procs': [], 'vsize': 0, 'rss': 0}
363 slice['procs'].append(proc)
364 slice['vsize'] += proc['vsize']
365 slice['rss'] += proc['rss']
367 slices[proc['xid']] = slice
373 Returns total physical memory on the system in KiB.
376 meminfo = open("/proc/meminfo", "r")
377 line = meminfo.readline()
379 if line[0:8] == "MemTotal":
380 # MemTotal: 255396 kB
381 (name, value, kb) = line.split()
388 Returns swap utilization on the system as a whole percentage (0-100).
395 swaps = open("/proc/swaps", "r")
397 lines = swaps.readlines()[1:]
400 # /dev/mapper/planetlab-swap partition 1048568 3740 -1
401 (filename, type, size, used, priority) = line.strip().split()
403 total_swap += int(size)
404 total_used += int(used)
405 except ValueEror, err:
407 except (IOError, KeyError), err:
410 return 100 * total_used / total_swap
412 def summary(names = None, total_rss = memtotal()):
414 Return a summary of memory usage by slice.
416 slicelist = slicestat(names).values()
417 slicelist.sort(lambda a, b: b['rss'] - a['rss'])
419 table = "%-20s%10s%24s\n\n" % ("Slice", "Processes", "Memory Usage")
420 for slice in slicelist:
421 table += "%-20s%10d%16s (%4.1f%%)\n" % \
422 (slice['name'], len(slice['procs']),
423 format_bytes(slice['rss'] * 1024, si = False),
424 100. * slice['rss'] / total_rss)
430 global debug, verbose, datafile
431 global period, change_thresh, reset_thresh, reboot_thresh, min_thresh, system_slices
436 longopts = ["debug", "verbose", "file=", "slice=", "status", "help"]
437 longopts += ["period=", "reset-thresh=", "reboot-thresh=", "min-thresh=", "system-slice="]
438 (opts, argv) = getopt.getopt(sys.argv[1:], "dvf:s:ph", longopts)
439 except getopt.GetoptError, err:
440 print "Error: " + err.msg
444 for (opt, optval) in opts:
445 if opt == "-d" or opt == "--debug":
447 elif opt == "-v" or opt == "--verbose":
449 elif opt == "-f" or opt == "--file":
451 elif opt == "-s" or opt == "--slice":
453 elif opt == "-p" or opt == "--period":
455 elif opt == "--change-thresh":
456 change_thresh = int(optval)
457 elif opt == "--reset-thresh":
458 reset_thresh = int(optval)
459 elif opt == "--reboot-thresh":
460 reboot_thresh = int(optval)
461 elif opt == "--min-thresh":
462 min_thresh = int(optval)
463 elif opt == "--system-slice":
464 system_slices.append(optval)
465 elif opt == "--status":
472 # Check if we are already running
479 # Redirect stdout and stderr to syslog
480 syslog.openlog("swapmon")
481 sys.stdout = sys.stderr = Logger()
483 # Get total physical memory
484 total_rss = memtotal()
487 f = open(datafile, "r+")
489 print "Loading %s" % datafile
490 (version, slices) = pickle.load(f)
492 # Check version of data file
493 if version != "$Id: swapmon.py,v 1.5 2006/05/09 03:23:57 mlhuang Exp $":
494 print "Not using old version '%s' data file %s" % (version, datafile)
497 params = {'hostname': socket.gethostname(),
498 'date': time.asctime(time.gmtime()) + " GMT",
499 'table': summary(total_rss)}
502 print rebooted_subject % params
503 print rebooted_body % params
505 slicemail(None, rebooted_subject % params, rebooted_body % params)
510 version = "$Id: swapmon.py,v 1.5 2006/05/09 03:23:57 mlhuang Exp $"
513 # Query process table every 30 seconds, or when a large change in
514 # swap utilization is detected.
519 # System slices that we have warned but could not reset
522 # Slices that were reset
528 for resetslice in resetlist.keys():
529 resetlist[resetslice].update()
531 if last_used is None:
534 print "%d%% swap consumed" % used
536 if used >= reboot_thresh:
537 # Dump slice state before rebooting
539 print "Saving %s" % datafile
540 f = open(datafile, "w")
541 pickle.dump((version, slices), f)
544 # Goodbye, cruel world
545 print "%d%% swap consumed, rebooting" % used
547 bwlimit.run("/bin/sync; /sbin/reboot -f")
549 elif used >= reset_thresh:
551 slicelist = slices.values()
552 slicelist.sort(lambda a, b: b['rss'] - a['rss'])
553 for slice in slicelist:
554 percent = 100. * slice['rss'] / total_rss
556 if percent < min_thresh:
559 print "%d%% swap consumed, slice %s is using %s (%d%%) of memory" % \
562 format_bytes(slice['rss'] * 1024, si = False),
565 slice['procs'].sort(lambda a, b: b['rss'] - a['rss'])
567 table = "%5s %10s %10s %4s %4s %s\n\n" % ("PID", "VIRT", "RES", '%CPU', '%MEM', 'COMMAND')
568 for proc in slice['procs']:
569 table += "%5s %10s %10s %4.1f %4.1f %s\n" % \
571 format_bytes(proc['vsize'] * 1024, si = False),
572 format_bytes(proc['rss'] * 1024, si = False),
573 proc['pcpu'], proc['pmem'], proc['cmd'])
575 params = {'hostname': socket.gethostname(),
576 'date': time.asctime(time.gmtime()) + " GMT",
578 'slice': slice['name'],
579 'rss': format_bytes(slice['rss'] * 1024, si = False),
582 # Match slice name against system slice patterns
583 is_system_slice = filter(None, [re.match(pattern, slice['name']) for pattern in system_slices])
586 if slice['name'] not in warned:
587 warned.append(slice['name'])
589 print alarm_subject % params
590 print alarm_body % params
592 print "Warning slice " + slice['name']
593 slicemail(slice['name'], alarm_subject % params,
597 if not resetlist.has_key(slice['name']):
598 resetlist[slice['name']] = Reset(slice['name'])
599 resetlist[slice['name']].reset(params)
601 elif timer <= 0 or used >= (last_used + change_thresh):
602 if used >= (last_used + change_thresh):
603 print "%d%% swap consumed, %d%% in last %d seconds" % \
604 (used, used - last_used, period - timer)
606 slices = slicestat(names)
609 # Keep track of large changes in swap utilization
617 if __name__ == '__main__':