#!/usr/bin/env python # Try to determine how much RAM is currently being used per program. # Note the per program, not per process. So for example this script # will report mem used by all httpd process together. In detail it reports: # sum(all RSS for process instances) + max(shared mem for any process instance) # # The shared calculation below will factor out shared text and # libs etc. within a program, but not between programs. So there # will always be some overestimation. This will be the same for # all processes that just use libc for e.g. but more for others # that use larger shared libs like gnome, kde etc. # Author: P@draigBrady.com # V1.0 06 Jul 2005 Initial release # V1.1 11 Aug 2006 root permission required for accuracy # V1.2 08 Nov 2006 Add total to output # Use KiB,MiB,... for units rather than K,M,... # V1.3 22 Nov 2006 Ignore shared col from /proc/$pid/statm for # 2.6 kernels up to and including 2.6.9. # There it represented the total file backed extent # V1.4 23 Nov 2006 Remove total from output as it's meaningless # (the shared values overlap with other programs). # Display the shared column. This extra info is # useful, especially as it overlaps between programs. # V1.5 26 Mar 2007 Remove redundant recursion from human() # V1.6 05 Jun 2007 Also report number of processes with a given name. # Patch from riccardo.murri@gmail.com # Notes: # # All interpreted programs where the interpreter is started # by the shell or with env, will be merged to the interpreter # (as that's what's given to exec). For e.g. all python programs # starting with "#!/usr/bin/env python" will be grouped under python. # You can change this by changing comm= to args= below but that will # have the undesirable affect of splitting up programs started with # differing parameters (for e.g. mingetty tty[1-6]). # # For 2.6 kernels up to and including 2.6.13 and later 2.4 redhat kernels # (rmap vm without smaps) it can not be accurately determined how many pages # are shared between processes in general or within a program in our case: # http://lkml.org/lkml/2005/7/6/250 # A warning is printed if overestimation is possible. # In addition for 2.6 kernels up to 2.6.9 inclusive, the shared # value in /proc/$pid/statm is the total file-backed extent of a process. # We ignore that, introducing more overestimation, again printing a warning. # # I don't take account of memory allocated for a program # by other programs. For e.g. memory used in the X server for # a program could be determined, but is not. # # This script assumes threads are already merged by ps # TODO: # # use ps just to enumerate the pids and names # so as to remove the race between reading rss and shared values import sys, os, string if os.geteuid() != 0: sys.stderr.write("Sorry, root permission required.\n"); sys.exit(1) PAGESIZE=os.sysconf("SC_PAGE_SIZE")/1024 #KiB our_pid=os.getpid() #(major,minor,release) def kernel_ver(): kv=open("/proc/sys/kernel/osrelease").readline().split(".")[:3] for char in "-_": kv[2]=kv[2].split(char)[0] return (int(kv[0]), int(kv[1]), int(kv[2])) kv=kernel_ver() def getShared(pid): if os.path.exists("/proc/"+str(pid)+"/smaps"): shared_lines=[line for line in open("/proc/"+str(pid)+"/smaps").readlines() if line.find("Shared")!=-1] return sum([int(line.split()[1]) for line in shared_lines]) elif (2,6,1) <= kv <= (2,6,9): return 0 #lots of overestimation, but what can we do? else: return int(open("/proc/"+str(pid)+"/statm").readline().split()[2])*PAGESIZE cmds={} shareds={} count={} for line in os.popen("ps -e -o rss=,pid=,comm=").readlines(): size, pid, cmd = map(string.strip,line.strip().split(None,2)) if int(pid) == our_pid: continue #no point counting this process try: shared=getShared(pid) except: continue #ps gone away if shareds.get(cmd): if shareds[cmd] < shared: shareds[cmd]=shared else: shareds[cmd]=shared #Note shared is always a subset of rss (trs is not always) cmds[cmd]=cmds.setdefault(cmd,0)+int(size)-shared if count.has_key(cmd): count[cmd] += 1 else: count[cmd] = 1 #Add max shared mem for each program for cmd in cmds.keys(): cmds[cmd]=cmds[cmd]+shareds[cmd] sort_list = cmds.items() sort_list.sort(lambda x,y:cmp(x[1],y[1])) sort_list=filter(lambda x:x[1],sort_list) #get rid of zero sized processes (kernel threads) #The following matches "du -h" output #see also human.py def human(num, power="Ki"): powers=["Ki","Mi","Gi","Ti"] while num >= 1000: #4 digits num /= 1024.0 power=powers[powers.index(power)+1] return "%.1f %s" % (num,power) def cmd_with_count(cmd, count): if count>1: return "%s (%u)" % (cmd, count) else: return cmd print " Private + Shared = RAM used\tProgram \n" for cmd in sort_list: print "%8sB + %8sB = %8sB\t%s" % (human(cmd[1]-shareds[cmd[0]]), human(shareds[cmd[0]]), human(cmd[1]), cmd_with_count(cmd[0], count[cmd[0]])) print "\n Private + Shared = RAM used\tProgram \n" #Warn of possible inaccuracies #1 = accurate #0 = some shared mem not reported #-1= all shared mem not reported def shared_val_accurate(): """http://wiki.apache.org/spamassassin/TopSharedMemoryBug""" if kv[:2] == (2,4): if open("/proc/meminfo").read().find("Inact_") == -1: return 1 return 0 elif kv[:2] == (2,6): if os.path.exists("/proc/"+str(os.getpid())+"/smaps"): return 1 if (2,6,1) <= kv <= (2,6,9): return -1 return 0 else: return 1 vm_accuracy = shared_val_accurate() if vm_accuracy == -1: sys.stderr.write("Warning: Shared memory is not reported by this system.\n") sys.stderr.write("Values reported will be too large.\n") elif vm_accuracy == 0: sys.stderr.write("Warning: Shared memory is not reported accurately by this system.\n") sys.stderr.write("Values reported could be too large.\n")