X-Git-Url: http://git.onelab.eu/?p=mom.git;a=blobdiff_plain;f=swapmon.py;h=e2215b279ed4a54fafa66cfee407e6101693b751;hp=35a98e86ba362f9318af3fcf278a801d9ddbe0e3;hb=35fc6f409f7372549b7c07bddecc6dc9541160c9;hpb=04914713a79dbac6bc7af8317fb978823ec743d0 diff --git a/swapmon.py b/swapmon.py index 35a98e8..e2215b2 100755 --- a/swapmon.py +++ b/swapmon.py @@ -22,10 +22,6 @@ import pickle import socket import time -# util-vserver/python/vserver.py allows us to control slices directly -# from Python -from vserver import VServer - # bwlimit exports a few useful functions like run(), get_xid(), and get_slice() import bwlimit @@ -50,15 +46,6 @@ reset_thresh = 80 # Swap utilization at which the machine is rebooted reboot_thresh = 95 -# 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 @@ -84,28 +71,6 @@ statistics are not entirely accurate due to threading. %(date)s %(hostname)s reboot """.lstrip() -# Message sent after a hog is reset -reset_subject = "pl_mom reset slice %(slice)s on %(hostname)s" -reset_body = \ -""" -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%%) (%(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 - -%(date)s %(hostname)s reset %(slice)s -""".lstrip() - # Message sent to system slices that should not be reset alarm_subject = "pl_mom alarm slice %(slice)s on %(hostname)s" alarm_body = \ @@ -145,111 +110,10 @@ and what you are doing to address the problem. %(date)s %(hostname)s reset %(slice)s """.lstrip() +def killsliverprocs(xid): + bwlimit.run("/usr/sbin/vkill -s 9 -c %s 0" % xid) -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]... @@ -514,6 +378,11 @@ def main(): global period, change_thresh, reset_thresh, reboot_thresh, rss_min, system_slices # All slices names = [] + timer = period + last_used = None + used = None + warned = [] + emailed = {} try: longopts = ["debug", "verbose", "file=", "slice=", "status", "help"] @@ -569,28 +438,11 @@ def main(): # Query process table every 30 seconds, or when a large change in # swap utilization is detected. - timer = period - last_used = None - used = None - - # 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 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 writedat(slices) @@ -630,11 +482,17 @@ def main(): alarm_body % params) else: # 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 not debug: + if emailed.get(slice['name'], (time.time() + email_timeout + 1)) > (time.time() + email_timeout): + slicemail(slice['name'], kill_subject % params, kill_body % params) + emailed[slice['name']] = time.time() + else: + print kill_subject % params + print kill_body % params + print "Killing procs in %s" % slice['name'] + killsliverprocs(slice['xid']) + + # wait period before 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" % \