perf scripts python: Add new compaction-times script
authorTony Jones <tonyj@suse.com>
Mon, 17 Aug 2015 19:48:52 +0000 (12:48 -0700)
committerArnaldo Carvalho de Melo <acme@redhat.com>
Fri, 28 Aug 2015 17:53:40 +0000 (14:53 -0300)
This patch creates a new script (compaction-times) to report time
spent in mm compaction. It is possible to report times in nanoseconds
(default) or microseconds (-u).

The option -p will break down results by process id, -pv will further
decompose by each compaction entry/exit.

For each compaction entry/exit what is reported is controlled by the
options:

  -t   report only timing
  -m   report migration stats
  -ms  report migration scanner stats
  -fs  report free scanner stats

The default is to report all.

Entries may be further filtered by pid, pid-range or comm (regex).

The script is useful when analysing workloads that compact memory. The
most common example will be THP allocations on systems with a lot of
uptime that has fragmented memory.

This is an example of using the script to analyse a thpscale from
mmtests which deliberately fragments memory and allocates THP in 4
separate threads

  # Recording step, one of the following;
  $ perf record -e 'compaction:mm_compaction_*' ./workload
  # or:
  $ perf script record compaction-times

  # Reporting: basic
  total: 2444505743ns migration: moved=357738 failed=39275
  free_scanner: scanned=2705578 isolated=387875
  migration_scanner: scanned=414426 isolated=397013

  # Reporting: Per task stall times
  $ perf script report compaction-times -- -t -p
  total: 2444505743ns
  6384[thpscale]: 740800017ns
  6385[thpscale]: 274119512ns
  6386[thpscale]: 832961337ns
  6383[thpscale]: 596624877ns

  # Reporting: Per-compaction attempts for task 6385
  $ perf script report compaction-times -- -m -pv 6385
  total: 274119512ns migration: moved=14893 failed=24285
  6385[thpscale]: 274119512ns migration: moved=14893 failed=24285
  6385[thpscale].1: 3033277ns migration: moved=511 failed=1
  6385[thpscale].2: 9592094ns migration: moved=1524 failed=12
  6385[thpscale].3: 2495587ns migration: moved=512 failed=0
  6385[thpscale].4: 2561766ns migration: moved=512 failed=0
  6385[thpscale].5: 2523521ns migration: moved=512 failed=0
  ..... output continues ...

Changes since v1:
- report stats for isolate_migratepages and isolate_freepages
  (Vlastimil Babka)
- refactor code to achieve above
- add help text
- output to stdout/stderr explicitly

Signed-off-by: Tony Jones <tonyj@suse.com>
Cc: Mel Gorman <mgorman@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Link: http://lkml.kernel.org/r/1439840932-8933-1-git-send-email-tonyj@suse.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
tools/perf/scripts/python/bin/compaction-times-record [new file with mode: 0644]
tools/perf/scripts/python/bin/compaction-times-report [new file with mode: 0644]
tools/perf/scripts/python/compaction-times.py [new file with mode: 0644]

diff --git a/tools/perf/scripts/python/bin/compaction-times-record b/tools/perf/scripts/python/bin/compaction-times-record
new file mode 100644 (file)
index 0000000..6edcd40
--- /dev/null
@@ -0,0 +1,2 @@
+#!/bin/bash
+perf record -e compaction:mm_compaction_begin -e compaction:mm_compaction_end -e compaction:mm_compaction_migratepages -e compaction:mm_compaction_isolate_migratepages -e compaction:mm_compaction_isolate_freepages $@
diff --git a/tools/perf/scripts/python/bin/compaction-times-report b/tools/perf/scripts/python/bin/compaction-times-report
new file mode 100644 (file)
index 0000000..3dc1389
--- /dev/null
@@ -0,0 +1,4 @@
+#!/bin/bash
+#description: display time taken by mm compaction
+#args: [-h] [-u] [-p|-pv] [-t | [-m] [-fs] [-ms]] [pid|pid-range|comm-regex]
+perf script -s "$PERF_EXEC_PATH"/scripts/python/compaction-times.py $@
diff --git a/tools/perf/scripts/python/compaction-times.py b/tools/perf/scripts/python/compaction-times.py
new file mode 100644 (file)
index 0000000..239cb05
--- /dev/null
@@ -0,0 +1,311 @@
+# report time spent in compaction
+# Licensed under the terms of the GNU GPL License version 2
+
+# testing:
+# 'echo 1 > /proc/sys/vm/compact_memory' to force compaction of all zones
+
+import os
+import sys
+import re
+
+import signal
+signal.signal(signal.SIGPIPE, signal.SIG_DFL)
+
+usage = "usage: perf script report compaction-times.py -- [-h] [-u] [-p|-pv] [-t | [-m] [-fs] [-ms]] [pid|pid-range|comm-regex]\n"
+
+class popt:
+       DISP_DFL = 0
+       DISP_PROC = 1
+       DISP_PROC_VERBOSE=2
+
+class topt:
+       DISP_TIME = 0
+       DISP_MIG = 1
+       DISP_ISOLFREE = 2
+       DISP_ISOLMIG = 4
+       DISP_ALL = 7
+
+class comm_filter:
+       def __init__(self, re):
+               self.re = re
+
+       def filter(self, pid, comm):
+               m = self.re.search(comm)
+               return m == None or m.group() == ""
+
+class pid_filter:
+       def __init__(self, low, high):
+               self.low = (0 if low == "" else int(low))
+               self.high = (0 if high == "" else int(high))
+
+       def filter(self, pid, comm):
+               return not (pid >= self.low and (self.high == 0 or pid <= self.high))
+
+def set_type(t):
+       global opt_disp
+       opt_disp = (t if opt_disp == topt.DISP_ALL else opt_disp|t)
+
+def ns(sec, nsec):
+       return (sec * 1000000000) + nsec
+
+def time(ns):
+       return "%dns" % ns if opt_ns else "%dus" % (round(ns, -3) / 1000)
+
+class pair:
+       def __init__(self, aval, bval, alabel = None, blabel = None):
+               self.alabel = alabel
+               self.blabel = blabel
+               self.aval = aval
+               self.bval = bval
+
+       def __add__(self, rhs):
+               self.aval += rhs.aval
+               self.bval += rhs.bval
+               return self
+
+       def __str__(self):
+               return "%s=%d %s=%d" % (self.alabel, self.aval, self.blabel, self.bval)
+
+class cnode:
+       def __init__(self, ns):
+               self.ns = ns
+               self.migrated = pair(0, 0, "moved", "failed")
+               self.fscan = pair(0,0, "scanned", "isolated")
+               self.mscan = pair(0,0, "scanned", "isolated")
+
+       def __add__(self, rhs):
+               self.ns += rhs.ns
+               self.migrated += rhs.migrated
+               self.fscan += rhs.fscan
+               self.mscan += rhs.mscan
+               return self
+
+       def __str__(self):
+               prev = 0
+               s = "%s " % time(self.ns)
+               if (opt_disp & topt.DISP_MIG):
+                       s += "migration: %s" % self.migrated
+                       prev = 1
+               if (opt_disp & topt.DISP_ISOLFREE):
+                       s += "%sfree_scanner: %s" % (" " if prev else "", self.fscan)
+                       prev = 1
+               if (opt_disp & topt.DISP_ISOLMIG):
+                       s += "%smigration_scanner: %s" % (" " if prev else "", self.mscan)
+               return s
+
+       def complete(self, secs, nsecs):
+               self.ns = ns(secs, nsecs) - self.ns
+
+       def increment(self, migrated, fscan, mscan):
+               if (migrated != None):
+                       self.migrated += migrated
+               if (fscan != None):
+                       self.fscan += fscan
+               if (mscan != None):
+                       self.mscan += mscan
+
+
+class chead:
+       heads = {}
+       val = cnode(0);
+       fobj = None
+
+       @classmethod
+       def add_filter(cls, filter):
+               cls.fobj = filter
+
+       @classmethod
+       def create_pending(cls, pid, comm, start_secs, start_nsecs):
+               filtered = 0
+               try:
+                       head = cls.heads[pid]
+                       filtered = head.is_filtered()
+               except KeyError:
+                       if cls.fobj != None:
+                               filtered = cls.fobj.filter(pid, comm)
+                       head = cls.heads[pid] = chead(comm, pid, filtered)
+
+               if not filtered:
+                       head.mark_pending(start_secs, start_nsecs)
+
+       @classmethod
+       def increment_pending(cls, pid, migrated, fscan, mscan):
+               head = cls.heads[pid]
+               if not head.is_filtered():
+                       if head.is_pending():
+                               head.do_increment(migrated, fscan, mscan)
+                       else:
+                               sys.stderr.write("missing start compaction event for pid %d\n" % pid)
+
+       @classmethod
+       def complete_pending(cls, pid, secs, nsecs):
+               head = cls.heads[pid]
+               if not head.is_filtered():
+                       if head.is_pending():
+                               head.make_complete(secs, nsecs)
+                       else:
+                               sys.stderr.write("missing start compaction event for pid %d\n" % pid)
+
+       @classmethod
+       def gen(cls):
+               if opt_proc != popt.DISP_DFL:
+                       for i in cls.heads:
+                               yield cls.heads[i]
+
+       @classmethod
+       def str(cls):
+               return cls.val
+
+       def __init__(self, comm, pid, filtered):
+               self.comm = comm
+               self.pid = pid
+               self.val = cnode(0)
+               self.pending = None
+               self.filtered = filtered
+               self.list = []
+
+       def __add__(self, rhs):
+               self.ns += rhs.ns
+               self.val += rhs.val
+               return self
+
+       def mark_pending(self, secs, nsecs):
+               self.pending = cnode(ns(secs, nsecs))
+
+       def do_increment(self, migrated, fscan, mscan):
+               self.pending.increment(migrated, fscan, mscan)
+
+       def make_complete(self, secs, nsecs):
+               self.pending.complete(secs, nsecs)
+               chead.val += self.pending
+
+               if opt_proc != popt.DISP_DFL:
+                       self.val += self.pending
+
+                       if opt_proc == popt.DISP_PROC_VERBOSE:
+                               self.list.append(self.pending)
+               self.pending = None
+
+       def enumerate(self):
+               if opt_proc == popt.DISP_PROC_VERBOSE and not self.is_filtered():
+                       for i, pelem in enumerate(self.list):
+                               sys.stdout.write("%d[%s].%d: %s\n" % (self.pid, self.comm, i+1, pelem))
+
+       def is_pending(self):
+               return self.pending != None
+
+       def is_filtered(self):
+               return self.filtered
+
+       def display(self):
+               if not self.is_filtered():
+                       sys.stdout.write("%d[%s]: %s\n" % (self.pid, self.comm, self.val))
+
+
+def trace_end():
+       sys.stdout.write("total: %s\n" % chead.str())
+       for i in chead.gen():
+               i.display(),
+               i.enumerate()
+
+def compaction__mm_compaction_migratepages(event_name, context, common_cpu,
+       common_secs, common_nsecs, common_pid, common_comm,
+       common_callchain, nr_migrated, nr_failed):
+
+       chead.increment_pending(common_pid,
+               pair(nr_migrated, nr_failed), None, None)
+
+def compaction__mm_compaction_isolate_freepages(event_name, context, common_cpu,
+        common_secs, common_nsecs, common_pid, common_comm,
+        common_callchain, start_pfn, end_pfn, nr_scanned, nr_taken):
+
+       chead.increment_pending(common_pid,
+               None, pair(nr_scanned, nr_taken), None)
+
+def compaction__mm_compaction_isolate_migratepages(event_name, context, common_cpu,
+        common_secs, common_nsecs, common_pid, common_comm,
+        common_callchain, start_pfn, end_pfn, nr_scanned, nr_taken):
+
+       chead.increment_pending(common_pid,
+               None, None, pair(nr_scanned, nr_taken))
+
+def compaction__mm_compaction_end(event_name, context, common_cpu,
+       common_secs, common_nsecs, common_pid, common_comm,
+       common_callchain, zone_start, migrate_start, free_start, zone_end,
+       sync, status):
+
+       chead.complete_pending(common_pid, common_secs, common_nsecs)
+
+def compaction__mm_compaction_begin(event_name, context, common_cpu,
+       common_secs, common_nsecs, common_pid, common_comm,
+       common_callchain, zone_start, migrate_start, free_start, zone_end,
+       sync):
+
+       chead.create_pending(common_pid, common_comm, common_secs, common_nsecs)
+
+def pr_help():
+       global usage
+
+       sys.stdout.write(usage)
+       sys.stdout.write("\n")
+       sys.stdout.write("-h    display this help\n")
+       sys.stdout.write("-p    display by process\n")
+       sys.stdout.write("-pv   display by process (verbose)\n")
+       sys.stdout.write("-t    display stall times only\n")
+       sys.stdout.write("-m    display stats for migration\n")
+       sys.stdout.write("-fs   display stats for free scanner\n")
+       sys.stdout.write("-ms   display stats for migration scanner\n")
+       sys.stdout.write("-u    display results in microseconds (default nanoseconds)\n")
+
+
+comm_re = None
+pid_re = None
+pid_regex = "^(\d*)-(\d*)$|^(\d*)$"
+
+opt_proc = popt.DISP_DFL
+opt_disp = topt.DISP_ALL
+
+opt_ns = True
+
+argc = len(sys.argv) - 1
+if argc >= 1:
+       pid_re = re.compile(pid_regex)
+
+       for i, opt in enumerate(sys.argv[1:]):
+               if opt[0] == "-":
+                       if opt == "-h":
+                               pr_help()
+                               exit(0);
+                       elif opt == "-p":
+                               opt_proc = popt.DISP_PROC
+                       elif opt == "-pv":
+                               opt_proc = popt.DISP_PROC_VERBOSE
+                       elif opt == '-u':
+                               opt_ns = False
+                       elif opt == "-t":
+                               set_type(topt.DISP_TIME)
+                       elif opt == "-m":
+                               set_type(topt.DISP_MIG)
+                       elif opt == "-fs":
+                               set_type(topt.DISP_ISOLFREE)
+                       elif opt == "-ms":
+                               set_type(topt.DISP_ISOLMIG)
+                       else:
+                               sys.exit(usage)
+
+               elif i == argc - 1:
+                       m = pid_re.search(opt)
+                       if m != None and m.group() != "":
+                               if m.group(3) != None:
+                                       f = pid_filter(m.group(3), m.group(3))
+                               else:
+                                       f = pid_filter(m.group(1), m.group(2))
+                       else:
+                               try:
+                                       comm_re=re.compile(opt)
+                               except:
+                                       sys.stderr.write("invalid regex '%s'" % opt)
+                                       sys.exit(usage)
+                               f = comm_filter(comm_re)
+
+                       chead.add_filter(f)