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 1.11 2006/12/02 19:11:47 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 # Physical size threshold to be considered a consumer. Rationale is if there are no procs
66 # with a size at least as large as this, then there is a slow leaker; better to just reboot.
69 # System slices that should not be reset (regexps)
70 system_slices = ['root', PLC_SLICE_PREFIX + '_']
72 # Message sent after a critical reboot
73 rebooted_subject = "pl_mom rebooted %(hostname)s"
76 Sometime before %(date)s, swap space was
77 nearly exhausted on %(hostname)s, so pl_mom rebooted it.
79 Slices active prior to reboot are listed below. Memory usage
80 statistics are not entirely accurate due to threading.
84 %(date)s %(hostname)s reboot
87 # Message sent after a hog is reset
88 reset_subject = "pl_mom reset slice %(slice)s on %(hostname)s"
91 Sometime before %(date)s, swap space was
92 nearly exhausted on %(hostname)s.
94 Slice %(slice)s was reset since it was the largest consumer of
95 physical memory at %(rss)s (%(percent)4.1f%%) (%(sz)s writable).
97 Please reply to this message explaining the nature of your experiment,
98 and what you are doing to address the problem.
100 %(slice)s processes prior to reset:
104 %(date)s %(hostname)s reset %(slice)s
107 # Message sent to system slices that should not be reset
108 alarm_subject = "pl_mom alarm slice %(slice)s on %(hostname)s"
111 Sometime before %(date)s, swap space was
112 nearly exhausted on %(hostname)s.
114 System slice %(slice)s was the largest consumer of physical memory at
115 %(rss)s (%(percent)4.1f%%) (%(sz)s writable). It was not reset,
116 but please verify its behavior.
118 %(slice)s processes prior to alarm:
122 %(date)s %(hostname)s alarm %(slice)s
125 # Message sent after a slice has been killed
126 kill_subject = "pl_mom killed slice %(slice)s on %(hostname)s"
129 Sometime before %(date)s, swap space was
130 nearly exhausted on %(hostname)s.
132 Slice %(slice)s was killed since it was the largest consumer of
133 physical memory at %(rss)s (%(percent)4.1f%%) (%(sz)s writable)
134 after repeated restarts.
136 Please reply to this message explaining the nature of your experiment,
137 and what you are doing to address the problem.
139 %(slice)s processes prior to reset:
143 %(date)s %(hostname)s reset %(slice)s
150 Keeps track of state information for resets and kills
152 resettimeleft - timeout before checking for next reset
153 resetcount - number of strikes
154 killtimeleft - time out before removing from kill queue
155 {kill,reset}mail - Time of last email
156 kill - State of kill. If slice is already being killed, wait before retry.
159 def __init__(self,name):
161 self.resettimeleft = reset_timeout
164 self.killtimeleft = kill_timeout
171 # Count down for next check of reset slice.
172 if self.resettimeleft > 0:
173 self.resettimeleft -= 1
174 if debug and verbose: 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
186 # Check to see if a slice needs to be killed. If it has been killed more
187 # than kill_thresh in the probation period (kill_timeout) send an email, kill the slice.
188 def checkkill(self,params):
189 if self.killtimeleft > 0 and self.resetcount >= kill_thresh:
191 print kill_subject % params
192 print kill_body % params
196 print "Slice %s is being killed." % self.name
197 vserver = VServer(self.name)
202 except Exception, err:
203 print "Warning: Exception received while killing slice %s: %s" \
205 if (time.time() - self.killmail) > email_timeout:
206 slicemail(self.name, kill_subject % params, kill_body % params)
207 print "Sending KILL email for slice %s" % self.name
208 self.killmail = time.time()
212 # Reset slice after checking to see if slice is out of timeout.
213 # Increment resetcount, check to see if larger than kill_thresh.
214 def reset(self, params):
215 # If its the first reset (came back after kill)
216 # or if its been reset before
217 # and we are out of the reset timeout.
218 if self.resetcount == 0 or self.resettimeleft == 0:
219 # Do we need to kill this slice? Check history first.
220 if self.checkkill(params): return
223 self.killtimeleft = kill_timeout
224 self.resettimeleft = reset_timeout
225 print "%s has %s seconds to die and has been reset %s times" \
226 %(self.name, self.resettimeleft, self.resetcount)
228 print reset_subject % params
229 print reset_body % params
233 print "Resetting slice " + self.name
234 vserver = VServer(self.name)
236 vserver.start(wait = False)
240 except Exception, err:
241 print "Warning: Exception received while resetting slice %s:" \
243 if (time.time() - self.resetmail) > email_timeout:
244 slicemail(self.name, reset_subject % params, reset_body % params)
245 print "Sending Reset email for slice %s" % self.name
246 self.resetmail = time.time()
250 Usage: %s [OPTIONS]...
253 -d, --debug Enable debugging (default: %s)
254 -v, --verbose Increase verbosity level (default: %d)
255 -f, --file=FILE Data file (default: %s)
256 -s, --slice=SLICE Constrain monitoring to these slices (default: all)
257 -p, --period=SECONDS Seconds between normal process analysis (default: %s)
258 --reset-thresh=PERCENT Swap utilization at which slice reset is attempted
259 --reboot-thresh=PERCENT Swap utilization at which the machine is rebooted
260 --min-thresh=PERCENT Minimum physical memory utilization to be considered a hog
261 --system-slice=SLICE System slice that should not be reset
262 --status Print memory usage statistics and exit
263 -h, --help This message
264 """.lstrip() % (sys.argv[0], debug, verbose, datafile, format_period(period))
266 def slicestat(names = None):
268 Get status of specified slices (if names is None or empty, all
269 slices). vsize, sz, and rss are in KiB. Returns
271 {xid: {'xid': slice_id,
273 'procs': [{'pid': pid, 'xid': slice_id, 'user', username, 'cmd': command,
274 'vsize': virtual_kib, 'sz': potential_kib, 'rss': physical_kib,
275 'pcpu': cpu_percent, 'pmem': mem_percent}]
276 'vsize': total_virtual_kib,
277 'sz': total_potential_kib,
278 'rss': total_physical_kib}}
281 # Mandatory fields. xid is a virtual field inserted by vps. Make
282 # sure cmd is last so that it does not get truncated
284 fields = ['pid', 'xid', 'user', 'vsize', 'sz', 'rss', 'pcpu', 'pmem', 'cmd']
286 # vps inserts xid after pid in the output, but ps doesn't know
287 # what the field means.
288 ps_fields = list(fields)
289 ps_fields.remove('xid')
293 # Eat the header line. vps depends on the header to figure out
294 # which column is the PID column, so we can't just tell ps not to
296 for line in bwlimit.run("/usr/sbin/vps -e -o " + ",".join(ps_fields))[1:]:
300 # Replace "0 MAIN" and "1 ALL_PROC" (the special monikers that
301 # vps uses to denote the root context and the "all contexts"
302 # context) with "0" so that we can just split() on whitespace.
303 line = line.replace("0 MAIN", "0").replace("1 ALL_PROC", "0")
305 # Represent process as a dict of fields
306 values = line.split(None, len(fields) - 1)
307 if len(values) != len(fields):
309 proc = dict(zip(fields, values))
311 # Convert ints and floats
314 proc[field] = int(proc[field])
317 proc[field] = float(proc[field])
321 # vps sometimes prints ERR instead of a context ID if it
322 # cannot identify the context of an orphaned (usually dying)
323 # process. Skip these processes.
324 if type(proc['xid']) != int:
327 # Assign (pl_)sshd processes to slice instead of root
328 m = re.search(r"sshd: ([a-zA-Z_]+)", proc['cmd'])
330 xid = bwlimit.get_xid(m.group(1))
334 name = bwlimit.get_slice(proc['xid'])
336 # Orphaned (not associated with a slice) class
337 name = "%d?" % proc['xid']
339 # Monitor only the specified slices
340 if names and name not in names:
343 # Additional overhead calculations from slicestat
345 # Include 12 KiB of process overhead =
346 # 4 KiB top-level page table +
347 # 4 KiB kernel structure +
348 # 4 KiB basic page table
351 # Include additional page table overhead
352 if proc['vsize'] > 4096:
353 proc['rss'] += 4 * ((proc['vsize'] - 1) / 4096)
355 if slices.has_key(proc['xid']):
356 slice = slices[proc['xid']]
358 slice = {'xid': proc['xid'], 'name': name, 'procs': [], 'vsize': 0, 'sz': 0, 'rss': 0}
360 slice['procs'].append(proc)
361 slice['vsize'] += proc['vsize']
362 slice['sz'] += proc['sz']
363 slice['rss'] += proc['rss']
365 slices[proc['xid']] = slice
371 Returns total physical and swap memory on the system in KiB.
377 meminfo = open("/proc/meminfo", "r")
378 for line in meminfo.readlines():
380 (name, value, kb) = line.split()
383 if name == "MemTotal:":
385 elif name == "SwapTotal:":
393 Returns swap utilization on the system as a whole percentage (0-100).
400 swaps = open("/proc/swaps", "r")
402 lines = swaps.readlines()[1:]
405 # /dev/mapper/planetlab-swap partition 1048568 3740 -1
406 (filename, type, size, used, priority) = line.strip().split()
408 total_swap += int(size)
409 total_used += int(used)
410 except ValueEror, err:
412 except (IOError, KeyError), err:
415 return 100 * total_used / total_swap
417 def summary(slices = None, total_mem = None, total_swap = None):
419 Return a summary of memory usage by slice.
423 slicelist = slices.values()
424 slicelist.sort(lambda a, b: b['sz'] - a['sz'])
425 if total_mem is None or total_swap is None:
426 (total_mem, total_swap) = memtotal()
428 table = "%-20s%10s%24s%24s\n\n" % ("Slice", "Processes", "Memory Usage", "Potential Usage")
429 for slice in slicelist:
430 table += "%-20s%10d%16s (%4.1f%%)%16s (%4.1f%%)\n" % \
431 (slice['name'], len(slice['procs']),
432 format_bytes(slice['rss'] * 1024, si = False),
433 100. * slice['rss'] / total_mem,
434 format_bytes(slice['sz'] * 1024, si = False),
435 100. * slice['sz'] / (total_mem + total_swap))
442 global debug, verbose, datafile
443 global period, change_thresh, reset_thresh, reboot_thresh, rss_min, system_slices
448 longopts = ["debug", "verbose", "file=", "slice=", "status", "help"]
449 longopts += ["period=", "reset-thresh=", "reboot-thresh=", "min-thresh=", "system-slice="]
450 (opts, argv) = getopt.getopt(sys.argv[1:], "dvf:s:ph", longopts)
451 except getopt.GetoptError, err:
452 print "Error: " + err.msg
456 for (opt, optval) in opts:
457 if opt == "-d" or opt == "--debug":
459 elif opt == "-v" or opt == "--verbose":
461 elif opt == "-f" or opt == "--file":
463 elif opt == "-s" or opt == "--slice":
465 elif opt == "-p" or opt == "--period":
467 elif opt == "--change-thresh":
468 change_thresh = int(optval)
469 elif opt == "--reset-thresh":
470 reset_thresh = int(optval)
471 elif opt == "--reboot-thresh":
472 reboot_thresh = int(optval)
473 elif opt == "--min-thresh":
474 rss_min = int(optval)
475 elif opt == "--system-slice":
476 system_slices.append(optval)
477 elif opt == "--status":
478 print summary(slicestat(names))
484 # Check if we are already running
491 # Redirect stdout and stderr to syslog
492 syslog.openlog("swapmon")
493 sys.stdout = sys.stderr = Logger()
496 (total_mem, total_swap) = memtotal()
499 f = open(datafile, "r+")
501 print "Loading %s" % datafile
502 (version, slices) = pickle.load(f)
504 # Check version of data file
505 if version != "$Id: swapmon.py,v 1.11 2006/12/02 19:11:47 mlhuang Exp $":
506 print "Not using old version '%s' data file %s" % (version, datafile)
509 params = {'hostname': socket.gethostname(),
510 'date': time.asctime(time.gmtime()) + " GMT",
511 'table': summary(slices, total_mem, total_swap)}
514 print rebooted_subject % params
515 print rebooted_body % params
517 slicemail(None, rebooted_subject % params, rebooted_body % params)
522 version = "$Id: swapmon.py,v 1.11 2006/12/02 19:11:47 mlhuang Exp $"
525 # Query process table every 30 seconds, or when a large change in
526 # swap utilization is detected.
531 # System slices that we have warned but could not reset
534 # Slices that were reset
540 for resetslice in resetlist.keys():
541 resetlist[resetslice].update()
543 if last_used is None:
547 print "%d%% swap consumed" % used
549 if used >= reboot_thresh:
550 # Dump slice state before rebooting
551 if verbose: print "Saving %s" % datafile
552 f = open(datafile, "w")
553 pickle.dump((version, slices), f)
556 # Goodbye, cruel world
557 print "%d%% swap consumed, rebooting" % used
558 if not debug: bwlimit.run("/bin/sync; /sbin/reboot -f")
560 elif used >= reset_thresh:
561 if debug: print "Memory used = %s" %(used)
563 slicelist = slices.values()
564 slicelist.sort(lambda a, b: b['rss'] - a['rss'])
565 for slice in slicelist:
566 percent = 100. * slice['rss'] / total_mem
567 if slice['rss'] < rss_min: continue
568 print "%d%% swap consumed, slice %s is using %s (%d%%) of memory" % \
571 format_bytes(slice['rss'] * 1024, si = False),
574 slice['procs'].sort(lambda a, b: b['rss'] - a['rss'])
575 table = "%5s %10s %10s %10s %4s %4s %s\n\n" % \
576 ("PID", "VIRT", "SZ", "RES", '%CPU', '%MEM', 'COMMAND')
577 for proc in slice['procs']:
578 table += "%5s %10s %10s %10s %4.1f %4.1f %s\n" % \
580 format_bytes(proc['vsize'] * 1024, si = False),
581 format_bytes(proc['sz'] * 1024, si = False),
582 format_bytes(proc['rss'] * 1024, si = False),
587 params = {'hostname': socket.gethostname(),
588 'date': time.asctime(time.gmtime()) + " GMT",
590 'slice': slice['name'],
591 'rss': format_bytes(slice['rss'] * 1024, si = False),
592 'sz': format_bytes(slice['sz'] * 1024, si = False),
595 # Match slice name against system slice patterns
596 is_system_slice = filter(None,
597 [re.match(pattern, slice['name']) for pattern in system_slices])
600 # Do not reset system slices, just warn once
601 if slice['name'] not in warned:
602 warned.append(slice['name'])
604 print alarm_subject % params
605 print alarm_body % params
607 print "Warning slice " + slice['name']
608 slicemail(slice['name'], alarm_subject % params,
612 if not resetlist.has_key(slice['name']):
613 resetlist[slice['name']] = Reset(slice['name'])
614 resetlist[slice['name']].reset(params)
615 slices = slicestat(names)
617 if timer <= 0 or used >= (last_used + change_thresh):
618 if used >= (last_used + change_thresh):
619 print "%d%% swap consumed, %d%% in last %d seconds" % \
620 (used, used - last_used, period - timer)
622 slices = slicestat(names)
625 # Keep track of large changes in swap utilization
633 if __name__ == '__main__':