X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=swapmon.py;h=b25d57f10b3b73336e249f146dbfddaf94de991c;hb=117f3864335ac652bfecb5efc33e95bc9a908741;hp=655ff476b8c905f2731828c5ab53aa1abb1ffcc0;hpb=b93c913a6355ebf6aaa4b19d25398cde34a21fc0;p=mom.git diff --git a/swapmon.py b/swapmon.py index 655ff47..b25d57f 100755 --- a/swapmon.py +++ b/swapmon.py @@ -7,9 +7,10 @@ # # Mark Huang # Andy Bavier +# Faiyaz Ahmed # Copyright (C) 2004-2006 The Trustees of Princeton University # -# $Id: swapmon.py,v 1.3 2006/05/01 18:28:01 mlhuang Exp $ +# $Id$ # import syslog @@ -34,8 +35,8 @@ from pl_mom import * # Defaults debug = False verbose = 0 -datafile = "/var/lib/misc/swapmon.dat" - +DATAFILE = "/var/lib/misc/swapmon.dat" +VERSION = "$Id$" # Seconds between process analysis period = 30 @@ -44,13 +45,26 @@ period = 30 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 = 25 + +# 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 + '_'] @@ -78,11 +92,13 @@ Sometime before %(date)s, swap space was nearly exhausted on %(hostname)s. Slice %(slice)s was reset since it was the largest consumer of -physical memory at %(rss)s (%(percent)4.1f%%). +physical memory at %(rss)s (%(percent)4.1f%%) (%(sz)s writable). Please reply to this message explaining the nature of your experiment, and what you are doing to address the problem. +http://summer.cs.princeton.edu/status/tabulator.cgi?table=slices/table_%(slice)s + %(slice)s processes prior to reset: %(table)s @@ -98,8 +114,8 @@ Sometime before %(date)s, swap space was nearly exhausted on %(hostname)s. System slice %(slice)s was the largest consumer of physical memory at -%(rss)s (%(percent)4.1f%%). It was not reset, but please verify its -behavior. +%(rss)s (%(percent)4.1f%%) (%(sz)s writable). It was not reset, +but please verify its behavior. %(slice)s processes prior to alarm: @@ -108,6 +124,132 @@ behavior. %(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%%) (%(sz)s writable) +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() + # ignore initscripts. Don't run anything at start. + vserver.INITSCRIPTS = [] + vserver.start() + 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() + 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]... @@ -124,26 +266,27 @@ Options: --system-slice=SLICE System slice that should not be reset --status Print memory usage statistics and exit -h, --help This message -""".lstrip() % (sys.argv[0], debug, verbose, datafile, format_period(period)) +""".lstrip() % (sys.argv[0], debug, verbose, DATAFILE, format_period(period)) def slicestat(names = None): """ Get status of specified slices (if names is None or empty, all - slices). vsize and rss are in KiB. Returns - + slices). vsize, sz, and rss are in KiB. Returns + PID CONTEXT VSZ SZ RSS %MEM CMD {xid: {'xid': slice_id, - 'name': slice_name, - 'procs': [{'pid': pid, 'xid': slice_id, 'user', username, 'cmd': command, - 'vsize': virtual_kib, 'rss': physical_kib, + 'name': slice_name, + 'procs': [{'pid': pid, 'xid': slice_id, 'cmd': command, + 'vsize': virtual_kib, 'sz': potential_kib, 'rss': physical_kib, 'pcpu': cpu_percent, 'pmem': mem_percent}] - 'vsize': total_virtual_kib, - 'rss': total_physical_kib}} + 'vsize': total_virtual_kib, + 'sz': total_potential_kib, + 'rss': total_physical_kib}} """ # Mandatory fields. xid is a virtual field inserted by vps. Make # sure cmd is last so that it does not get truncated # automatically. - fields = ['pid', 'xid', 'user', 'vsize', 'rss', 'pcpu', 'pmem', 'cmd'] + fields = ['pid', 'xid', 'vsize', 'sz', 'rss', 'pmem', 'cmd'] # vps inserts xid after pid in the output, but ps doesn't know # what the field means. @@ -155,7 +298,7 @@ def slicestat(names = None): # Eat the header line. vps depends on the header to figure out # which column is the PID column, so we can't just tell ps not to # print it. - for line in bwlimit.run("/usr/sbin/vps -e -o " + ",".join(ps_fields))[1:]: + for line in bwlimit.run("/usr/sbin/vps -e -o " + ":16,".join(ps_fields))[1:]: # Chomp newline line = line.strip() @@ -180,21 +323,25 @@ def slicestat(names = None): except ValueError: pass + # vps sometimes prints ERR or the name of the slice + # 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) or (type(proc['vsize']) !=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: xid = bwlimit.get_xid(m.group(1)) if xid is not None: proc['xid'] = xid - try: - name = bwlimit.get_slice(proc['xid']) - if name is None: - # Orphaned (not associated with a slice) class - name = "%d?" % proc['xid'] - except Exception, err: - print "Warning: Exception received while parsing vps output", err - print proc + name = bwlimit.get_slice(proc['xid']) + if name is None: + # Orphaned (not associated with a slice) class + name = "%d?" % proc['xid'] # Monitor only the specified slices if names and name not in names: @@ -215,39 +362,42 @@ def slicestat(names = None): if slices.has_key(proc['xid']): slice = slices[proc['xid']] else: - slice = {'xid': proc['xid'], 'name': name, 'procs': [], 'vsize': 0, 'rss': 0} + slice = {'xid': proc['xid'], 'name': name, 'procs': [], 'vsize': 0, 'sz': 0, 'rss': 0} slice['procs'].append(proc) slice['vsize'] += proc['vsize'] + slice['sz'] += proc['sz'] slice['rss'] += proc['rss'] slices[proc['xid']] = slice - + return slices def memtotal(): """ - Returns total physical memory on the system in KiB. + Returns total physical and swap memory on the system in KiB. """ - + mem = 0 + swap = 0 meminfo = open("/proc/meminfo", "r") - line = meminfo.readline() + for line in meminfo.readlines(): + try: + (name, value, kb) = line.split() + except: + continue + if name == "MemTotal:": + mem = int(value) + elif name == "SwapTotal:": + swap = int(value) meminfo.close() - if line[0:8] == "MemTotal": - # MemTotal: 255396 kB - (name, value, kb) = line.split() - return int(value) - - return 0 + return (mem, swap) def swap_used(): """ Returns swap utilization on the system as a whole percentage (0-100). """ - total_swap = 0 total_used = 0 - try: swaps = open("/proc/swaps", "r") # Eat header line @@ -261,31 +411,105 @@ def swap_used(): total_used += int(used) except ValueEror, err: pass - except (IOError, KeyError), err: - pass + except (IOError, KeyError), err: pass - return 100 * total_used / total_swap + swapused = 100 * total_used / total_swap + if debug: print "%s percent swap used" % swapused + return swapused -def summary(names = None, total_rss = memtotal()): +def summary(slices = None, total_mem = None, total_swap = None): """ Return a summary of memory usage by slice. """ - slicelist = slicestat(names).values() - slicelist.sort(lambda a, b: b['rss'] - a['rss']) + if not slices: slices = slicestat() + slicelist = slices.values() + slicelist.sort(lambda a, b: b['sz'] - a['sz']) + if total_mem is None or total_swap is None: + (total_mem, total_swap) = memtotal() - table = "%-20s%10s%24s\n\n" % ("Slice", "Processes", "Memory Usage") + table = "%-20s%10s%24s%24s\n\n" % ("Slice", "Processes", "Memory Usage", "Potential Usage") for slice in slicelist: - table += "%-20s%10d%16s (%4.1f%%)\n" % \ + table += "%-20s%10d%16s (%4.1f%%)%16s (%4.1f%%)\n" % \ (slice['name'], len(slice['procs']), format_bytes(slice['rss'] * 1024, si = False), - 100. * slice['rss'] / total_rss) - + 100. * slice['rss'] / total_mem, + format_bytes(slice['sz'] * 1024, si = False), + 100. * slice['sz'] / (total_mem + total_swap)) return table +def formtable(slice, percent): + ''' + Makes pretty message to email with human readable ps values. + ''' + table = "%5s %10s %10s %10s %4s %4s %s\n\n" % \ + ("PID", "VIRT", "SZ", "RES", '%CPU', '%MEM', 'COMMAND') + for proc in slice['procs']: + table += "%5s %10s %10s %10s %4.1f %s\n" % \ + (proc['pid'], + format_bytes(proc['vsize'] * 1024, si = False), + format_bytes(proc['sz'] * 1024, si = False), + format_bytes(proc['rss'] * 1024, si = False), + proc['pmem'], + proc['cmd']) + + prettytable = {'hostname': socket.gethostname(), + 'date': time.asctime(time.gmtime()) + " GMT", + 'table': table, + 'slice': slice['name'], + 'rss': format_bytes(slice['rss'] * 1024, si = False), + 'sz': format_bytes(slice['sz'] * 1024, si = False), + 'percent': percent} + return prettytable + +def readdat(): + ''' + Return dictionary of vps (slicestat) from datfile left behind by OOM + before rebooting. If none file, just grab the latest dict (slicestat) + and return that. If dat file found, means we rebooted, send an email to + pl_mom@pl. + ''' + try: + f = open(DATAFILE, "r+") + if verbose: + print "Loading %s" % DATAFILE + (v, slices) = pickle.load(f) + f.close() + # Check version of data file + if v != VERSION: + print "Not using old version '%s' data file %s" % (v, DATAFILE) + raise Exception + + params = {'hostname': socket.gethostname(), + 'date': time.asctime(time.gmtime()) + " GMT", + 'table': summary(slices, total_mem, total_swap)} + if debug: + print rebooted_subject % params + print rebooted_body % params + else: + slicemail(None, rebooted_subject % params, rebooted_body % params) + + # Delete data file + os.unlink(DATAFILE) + except Exception: + slices = slicestat() + + return slices + + +def writedat(slices): + """ + Write (slices) to pickled datfile. + """ + if verbose: print "Saving %s" % DATAFILE + f = open(DATAFILE, "w") + pickle.dump((VERSION, slices), f) + f.close() + + def main(): # Defaults - global debug, verbose, datafile - global period, change_thresh, reset_thresh, reboot_thresh, min_thresh, system_slices + global debug, verbose, DATAFILE, VERSION + global period, change_thresh, reset_thresh, reboot_thresh, rss_min, system_slices # All slices names = [] @@ -304,7 +528,7 @@ def main(): elif opt == "-v" or opt == "--verbose": verbose += 1 elif opt == "-f" or opt == "--file": - datafile = optval + DATAFILE = optval elif opt == "-s" or opt == "--slice": names.append(optval) elif opt == "-p" or opt == "--period": @@ -316,11 +540,11 @@ def main(): elif opt == "--reboot-thresh": reboot_thresh = int(optval) elif opt == "--min-thresh": - min_thresh = int(optval) + rss_min = int(optval) elif opt == "--system-slice": system_slices.append(optval) elif opt == "--status": - print summary(names) + print summary(slicestat(names)) sys.exit(0) else: usage() @@ -337,35 +561,9 @@ def main(): syslog.openlog("swapmon") sys.stdout = sys.stderr = Logger() - # Get total physical memory - total_rss = memtotal() - - try: - f = open(datafile, "r+") - if verbose: - print "Loading %s" % datafile - (version, slices) = pickle.load(f) - f.close() - # Check version of data file - if version != "$Id: swapmon.py,v 1.3 2006/05/01 18:28:01 mlhuang Exp $": - print "Not using old version '%s' data file %s" % (version, datafile) - raise Exception - - params = {'hostname': socket.gethostname(), - 'date': time.asctime(time.gmtime()) + " GMT", - 'table': summary(total_rss)} - - if debug: - print rebooted_subject % params - print rebooted_body % params - else: - slicemail(None, rebooted_subject % params, rebooted_body % params) - - # Delete data file - os.unlink(datafile) - except Exception: - version = "$Id: swapmon.py,v 1.3 2006/05/01 18:28:01 mlhuang Exp $" - slices = {} + # Get total memory + (total_mem, total_swap) = memtotal() + slices = readdat() # Query process table every 30 seconds, or when a large change in # swap utilization is detected. @@ -376,103 +574,75 @@ def main(): # System slices that we have warned but could not reset warned = [] + # Slices that were reset + resetlist = {} + while True: used = swap_used() - if last_used is None: - last_used = used - if verbose: - print "%d%% swap consumed" % used - + if last_used is None: last_used = used + + # If we've reset you recently, update timers. + for resetslice in resetlist.keys(): + resetlist[resetslice].update() + # If you've been good, remove you from our list. + if resetlist[resetslice].killtimeleft == 0 and \ + resetlist[resetslice].resettimeleft == 0: + del resetlist[resetslice] + if used >= reboot_thresh: # Dump slice state before rebooting - if verbose: - print "Saving %s" % datafile - f = open(datafile, "w") - pickle.dump((version, slices), f) - f.close() - + writedat(slices) # Goodbye, cruel world print "%d%% swap consumed, rebooting" % used - if not debug: - bwlimit.run("/bin/sync; /sbin/reboot -f") - + if not debug: bwlimit.run("/bin/sync; /sbin/reboot -f") elif used >= reset_thresh: # Try and find a hog slicelist = slices.values() + # Puts largest on top. slicelist.sort(lambda a, b: b['rss'] - a['rss']) for slice in slicelist: - percent = 100. * slice['rss'] / total_rss - if percent < min_thresh: - continue - + percent = 100. * slice['rss'] / total_mem + if slice['rss'] < rss_min: continue print "%d%% swap consumed, slice %s is using %s (%d%%) of memory" % \ - (used, - slice['name'], - format_bytes(slice['rss'] * 1024, si = False), - percent) - + (used, + slice['name'], + format_bytes(slice['rss'] * 1024, si = False), + percent) slice['procs'].sort(lambda a, b: b['rss'] - a['rss']) - - table = "%5s %10s %10s %4s %4s %s\n\n" % ("PID", "VIRT", "RES", '%CPU', '%MEM', 'COMMAND') - for proc in slice['procs']: - table += "%5s %10s %10s %4.1f %4.1f %s\n" % \ - (proc['pid'], - format_bytes(proc['vsize'] * 1024, si = False), - format_bytes(proc['rss'] * 1024, si = False), - proc['pcpu'], proc['pmem'], proc['cmd']) - - params = {'hostname': socket.gethostname(), - 'date': time.asctime(time.gmtime()) + " GMT", - 'table': table, - 'slice': slice['name'], - 'rss': format_bytes(slice['rss'] * 1024, si = False), - 'percent': percent} - + # Make a pretty table. + params = formtable(slice, percent) # 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 + is_system_slice = filter(None, + [re.match(pattern, slice['name']) for pattern in system_slices]) + + # Do not reset system slices, just warn once + if is_system_slice: if slice['name'] not in warned: warned.append(slice['name']) + print "Warning slice " + 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) + 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) + + # wait period vefore recalculating swap. If in danger, recalc. + 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) + (used, used - last_used, period - timer) # Get slice state slices = slicestat(names) # Reset timer timer = period # Keep track of large changes in swap utilization last_used = used - timer -= 1 time.sleep(1)