projects/linux·blob master · refs

989 lines · 30.0 KiB

1#!/usr/bin/env python3
2# SPDX-License-Identifier: GPL-2.0
3#
4# Run a perf script command multiple times in parallel, using perf script
5# options --cpu and --time so that each job processes a different chunk
6# of the data.
7#
8# Copyright (c) 2024, Intel Corporation.
9
10import subprocess
11import argparse
12import pathlib
13import shlex
14import time
15import copy
16import sys
17import os
18import re
19
20glb_prog_name = "parallel-perf.py"
21glb_min_interval = 10.0
22glb_min_samples = 64
23
24class Verbosity():
25
26 def __init__(self, quiet=False, verbose=False, debug=False):
27 self.normal = True
28 self.verbose = verbose
29 self.debug = debug
30 self.self_test = True
31 if self.debug:
32 self.verbose = True
33 if self.verbose:
34 quiet = False
35 if quiet:
36 self.normal = False
37
38# Manage work (Start/Wait/Kill), as represented by a subprocess.Popen command
39class Work():
40
41 def __init__(self, cmd, pipe_to, output_dir="."):
42 self.popen = None
43 self.consumer = None
44 self.cmd = cmd
45 self.pipe_to = pipe_to
46 self.output_dir = output_dir
47 self.cmdout_name = f"{output_dir}/cmd.txt"
48 self.stdout_name = f"{output_dir}/out.txt"
49 self.stderr_name = f"{output_dir}/err.txt"
50
51 def Command(self):
52 sh_cmd = [ shlex.quote(x) for x in self.cmd ]
53 return " ".join(self.cmd)
54
55 def Stdout(self):
56 return open(self.stdout_name, "w")
57
58 def Stderr(self):
59 return open(self.stderr_name, "w")
60
61 def CreateOutputDir(self):
62 pathlib.Path(self.output_dir).mkdir(parents=True, exist_ok=True)
63
64 def Start(self):
65 if self.popen:
66 return
67 self.CreateOutputDir()
68 with open(self.cmdout_name, "w") as f:
69 f.write(self.Command())
70 f.write("\n")
71 stdout = self.Stdout()
72 stderr = self.Stderr()
73 if self.pipe_to:
74 self.popen = subprocess.Popen(self.cmd, stdout=subprocess.PIPE, stderr=stderr)
75 args = shlex.split(self.pipe_to)
76 self.consumer = subprocess.Popen(args, stdin=self.popen.stdout, stdout=stdout, stderr=stderr)
77 else:
78 self.popen = subprocess.Popen(self.cmd, stdout=stdout, stderr=stderr)
79
80 def RemoveEmptyErrFile(self):
81 if os.path.exists(self.stderr_name):
82 if os.path.getsize(self.stderr_name) == 0:
83 os.unlink(self.stderr_name)
84
85 def Errors(self):
86 if os.path.exists(self.stderr_name):
87 if os.path.getsize(self.stderr_name) != 0:
88 return [ f"Non-empty error file {self.stderr_name}" ]
89 return []
90
91 def TidyUp(self):
92 self.RemoveEmptyErrFile()
93
94 def RawPollWait(self, p, wait):
95 if wait:
96 return p.wait()
97 return p.poll()
98
99 def Poll(self, wait=False):
100 if not self.popen:
101 return None
102 result = self.RawPollWait(self.popen, wait)
103 if self.consumer:
104 res = result
105 result = self.RawPollWait(self.consumer, wait)
106 if result != None and res == None:
107 self.popen.kill()
108 result = None
109 elif result == 0 and res != None and res != 0:
110 result = res
111 if result != None:
112 self.TidyUp()
113 return result
114
115 def Wait(self):
116 return self.Poll(wait=True)
117
118 def Kill(self):
119 if not self.popen:
120 return
121 self.popen.kill()
122 if self.consumer:
123 self.consumer.kill()
124
125def KillWork(worklist, verbosity):
126 for w in worklist:
127 w.Kill()
128 for w in worklist:
129 w.Wait()
130
131def NumberOfCPUs():
132 return os.sysconf("SC_NPROCESSORS_ONLN")
133
134def NanoSecsToSecsStr(x):
135 if x == None:
136 return ""
137 x = str(x)
138 if len(x) < 10:
139 x = "0" * (10 - len(x)) + x
140 return x[:len(x) - 9] + "." + x[-9:]
141
142def InsertOptionAfter(cmd, option, after):
143 try:
144 pos = cmd.index(after)
145 cmd.insert(pos + 1, option)
146 except:
147 cmd.append(option)
148
149def CreateWorkList(cmd, pipe_to, output_dir, cpus, time_ranges_by_cpu):
150 max_len = len(str(cpus[-1]))
151 cpu_dir_fmt = f"cpu-%.{max_len}u"
152 worklist = []
153 pos = 0
154 for cpu in cpus:
155 if cpu >= 0:
156 cpu_dir = os.path.join(output_dir, cpu_dir_fmt % cpu)
157 cpu_option = f"--cpu={cpu}"
158 else:
159 cpu_dir = output_dir
160 cpu_option = None
161
162 tr_dir_fmt = "time-range"
163
164 if len(time_ranges_by_cpu) > 1:
165 time_ranges = time_ranges_by_cpu[pos]
166 tr_dir_fmt += f"-{pos}"
167 pos += 1
168 else:
169 time_ranges = time_ranges_by_cpu[0]
170
171 max_len = len(str(len(time_ranges)))
172 tr_dir_fmt += f"-%.{max_len}u"
173
174 i = 0
175 for r in time_ranges:
176 if r == [None, None]:
177 time_option = None
178 work_output_dir = cpu_dir
179 else:
180 time_option = "--time=" + NanoSecsToSecsStr(r[0]) + "," + NanoSecsToSecsStr(r[1])
181 work_output_dir = os.path.join(cpu_dir, tr_dir_fmt % i)
182 i += 1
183 work_cmd = list(cmd)
184 if time_option != None:
185 InsertOptionAfter(work_cmd, time_option, "script")
186 if cpu_option != None:
187 InsertOptionAfter(work_cmd, cpu_option, "script")
188 w = Work(work_cmd, pipe_to, work_output_dir)
189 worklist.append(w)
190 return worklist
191
192def DoRunWork(worklist, nr_jobs, verbosity):
193 nr_to_do = len(worklist)
194 not_started = list(worklist)
195 running = []
196 done = []
197 chg = False
198 while True:
199 nr_done = len(done)
200 if chg and verbosity.normal:
201 nr_run = len(running)
202 print(f"\rThere are {nr_to_do} jobs: {nr_done} completed, {nr_run} running", flush=True, end=" ")
203 if verbosity.verbose:
204 print()
205 chg = False
206 if nr_done == nr_to_do:
207 break
208 while len(running) < nr_jobs and len(not_started):
209 w = not_started.pop(0)
210 running.append(w)
211 if verbosity.verbose:
212 print("Starting:", w.Command())
213 w.Start()
214 chg = True
215 if len(running):
216 time.sleep(0.1)
217 finished = []
218 not_finished = []
219 while len(running):
220 w = running.pop(0)
221 r = w.Poll()
222 if r == None:
223 not_finished.append(w)
224 continue
225 if r == 0:
226 if verbosity.verbose:
227 print("Finished:", w.Command())
228 finished.append(w)
229 chg = True
230 continue
231 if verbosity.normal and not verbosity.verbose:
232 print()
233 print("Job failed!\n return code:", r, "\n command: ", w.Command())
234 if w.pipe_to:
235 print(" piped to: ", w.pipe_to)
236 print("Killing outstanding jobs")
237 KillWork(not_finished, verbosity)
238 KillWork(running, verbosity)
239 return False
240 running = not_finished
241 done += finished
242 errorlist = []
243 for w in worklist:
244 errorlist += w.Errors()
245 if len(errorlist):
246 print("Errors:")
247 for e in errorlist:
248 print(e)
249 elif verbosity.normal:
250 print("\r"," "*50, "\rAll jobs finished successfully", flush=True)
251 return True
252
253def RunWork(worklist, nr_jobs=NumberOfCPUs(), verbosity=Verbosity()):
254 try:
255 return DoRunWork(worklist, nr_jobs, verbosity)
256 except:
257 for w in worklist:
258 w.Kill()
259 raise
260 return True
261
262def ReadHeader(perf, file_name):
263 return subprocess.Popen([perf, "script", "--header-only", "--input", file_name], stdout=subprocess.PIPE).stdout.read().decode("utf-8")
264
265def ParseHeader(hdr):
266 result = {}
267 lines = hdr.split("\n")
268 for line in lines:
269 if ":" in line and line[0] == "#":
270 pos = line.index(":")
271 name = line[1:pos-1].strip()
272 value = line[pos+1:].strip()
273 if name in result:
274 orig_name = name
275 nr = 2
276 while True:
277 name = f"{orig_name} {nr}"
278 if name not in result:
279 break
280 nr += 1
281 result[name] = value
282 return result
283
284def HeaderField(hdr_dict, hdr_fld):
285 if hdr_fld not in hdr_dict:
286 raise Exception(f"'{hdr_fld}' missing from header information")
287 return hdr_dict[hdr_fld]
288
289# Represent the position of an option within a command string
290# and provide the option value and/or remove the option
291class OptPos():
292
293 def Init(self, opt_element=-1, value_element=-1, opt_pos=-1, value_pos=-1, error=None):
294 self.opt_element = opt_element # list element that contains option
295 self.value_element = value_element # list element that contains option value
296 self.opt_pos = opt_pos # string position of option
297 self.value_pos = value_pos # string position of value
298 self.error = error # error message string
299
300 def __init__(self, args, short_name, long_name, default=None):
301 self.args = list(args)
302 self.default = default
303 n = 2 + len(long_name)
304 m = len(short_name)
305 pos = -1
306 for opt in args:
307 pos += 1
308 if m and opt[:2] == f"-{short_name}":
309 if len(opt) == 2:
310 if pos + 1 < len(args):
311 self.Init(pos, pos + 1, 0, 0)
312 else:
313 self.Init(error = f"-{short_name} option missing value")
314 else:
315 self.Init(pos, pos, 0, 2)
316 return
317 if opt[:n] == f"--{long_name}":
318 if len(opt) == n:
319 if pos + 1 < len(args):
320 self.Init(pos, pos + 1, 0, 0)
321 else:
322 self.Init(error = f"--{long_name} option missing value")
323 elif opt[n] == "=":
324 self.Init(pos, pos, 0, n + 1)
325 else:
326 self.Init(error = f"--{long_name} option expected '='")
327 return
328 if m and opt[:1] == "-" and opt[:2] != "--" and short_name in opt:
329 ipos = opt.index(short_name)
330 if "-" in opt[1:]:
331 hpos = opt[1:].index("-")
332 if hpos < ipos:
333 continue
334 if ipos + 1 == len(opt):
335 if pos + 1 < len(args):
336 self.Init(pos, pos + 1, ipos, 0)
337 else:
338 self.Init(error = f"-{short_name} option missing value")
339 else:
340 self.Init(pos, pos, ipos, ipos + 1)
341 return
342 self.Init()
343
344 def Value(self):
345 if self.opt_element >= 0:
346 if self.opt_element != self.value_element:
347 return self.args[self.value_element]
348 else:
349 return self.args[self.value_element][self.value_pos:]
350 return self.default
351
352 def Remove(self, args):
353 if self.opt_element == -1:
354 return
355 if self.opt_element != self.value_element:
356 del args[self.value_element]
357 if self.opt_pos:
358 args[self.opt_element] = args[self.opt_element][:self.opt_pos]
359 else:
360 del args[self.opt_element]
361
362def DetermineInputFileName(cmd):
363 p = OptPos(cmd, "i", "input", "perf.data")
364 if p.error:
365 raise Exception(f"perf command {p.error}")
366 file_name = p.Value()
367 if not os.path.exists(file_name):
368 raise Exception(f"perf command input file '{file_name}' not found")
369 return file_name
370
371def ReadOption(args, short_name, long_name, err_prefix, remove=False):
372 p = OptPos(args, short_name, long_name)
373 if p.error:
374 raise Exception(f"{err_prefix}{p.error}")
375 value = p.Value()
376 if remove:
377 p.Remove(args)
378 return value
379
380def ExtractOption(args, short_name, long_name, err_prefix):
381 return ReadOption(args, short_name, long_name, err_prefix, True)
382
383def ReadPerfOption(args, short_name, long_name):
384 return ReadOption(args, short_name, long_name, "perf command ")
385
386def ExtractPerfOption(args, short_name, long_name):
387 return ExtractOption(args, short_name, long_name, "perf command ")
388
389def PerfDoubleQuickCommands(cmd, file_name):
390 cpu_str = ReadPerfOption(cmd, "C", "cpu")
391 time_str = ReadPerfOption(cmd, "", "time")
392 # Use double-quick sampling to determine trace data density
393 times_cmd = ["perf", "script", "--ns", "--input", file_name, "--itrace=qqi"]
394 if cpu_str != None and cpu_str != "":
395 times_cmd.append(f"--cpu={cpu_str}")
396 if time_str != None and time_str != "":
397 times_cmd.append(f"--time={time_str}")
398 cnts_cmd = list(times_cmd)
399 cnts_cmd.append("-Fcpu")
400 times_cmd.append("-Fcpu,time")
401 return cnts_cmd, times_cmd
402
403class CPUTimeRange():
404 def __init__(self, cpu):
405 self.cpu = cpu
406 self.sample_cnt = 0
407 self.time_ranges = None
408 self.interval = 0
409 self.interval_remaining = 0
410 self.remaining = 0
411 self.tr_pos = 0
412
413def CalcTimeRangesByCPU(line, cpu, cpu_time_ranges, max_time):
414 cpu_time_range = cpu_time_ranges[cpu]
415 cpu_time_range.remaining -= 1
416 cpu_time_range.interval_remaining -= 1
417 if cpu_time_range.remaining == 0:
418 cpu_time_range.time_ranges[cpu_time_range.tr_pos][1] = max_time
419 return
420 if cpu_time_range.interval_remaining == 0:
421 time = TimeVal(line[1][:-1], 0)
422 time_ranges = cpu_time_range.time_ranges
423 time_ranges[cpu_time_range.tr_pos][1] = time - 1
424 time_ranges.append([time, max_time])
425 cpu_time_range.tr_pos += 1
426 cpu_time_range.interval_remaining = cpu_time_range.interval
427
428def CountSamplesByCPU(line, cpu, cpu_time_ranges):
429 try:
430 cpu_time_ranges[cpu].sample_cnt += 1
431 except:
432 print("exception")
433 print("cpu", cpu)
434 print("len(cpu_time_ranges)", len(cpu_time_ranges))
435 raise
436
437def ProcessCommandOutputLines(cmd, per_cpu, fn, *x):
438 # Assume CPU number is at beginning of line and enclosed by []
439 pat = re.compile(r"\s*\[[0-9]+\]")
440 p = subprocess.Popen(cmd, stdout=subprocess.PIPE)
441 while True:
442 line = p.stdout.readline()
443 if line:
444 line = line.decode("utf-8")
445 if pat.match(line):
446 line = line.split()
447 if per_cpu:
448 # Assumes CPU number is enclosed by []
449 cpu = int(line[0][1:-1])
450 else:
451 cpu = 0
452 fn(line, cpu, *x)
453 else:
454 break
455 p.wait()
456
457def IntersectTimeRanges(new_time_ranges, time_ranges):
458 pos = 0
459 new_pos = 0
460 # Can assume len(time_ranges) != 0 and len(new_time_ranges) != 0
461 # Note also, there *must* be at least one intersection.
462 while pos < len(time_ranges) and new_pos < len(new_time_ranges):
463 # new end < old start => no intersection, remove new
464 if new_time_ranges[new_pos][1] < time_ranges[pos][0]:
465 del new_time_ranges[new_pos]
466 continue
467 # new start > old end => no intersection, check next
468 if new_time_ranges[new_pos][0] > time_ranges[pos][1]:
469 pos += 1
470 if pos < len(time_ranges):
471 continue
472 # no next, so remove remaining
473 while new_pos < len(new_time_ranges):
474 del new_time_ranges[new_pos]
475 return
476 # Found an intersection
477 # new start < old start => adjust new start = old start
478 if new_time_ranges[new_pos][0] < time_ranges[pos][0]:
479 new_time_ranges[new_pos][0] = time_ranges[pos][0]
480 # new end > old end => keep the overlap, insert the remainder
481 if new_time_ranges[new_pos][1] > time_ranges[pos][1]:
482 r = [ time_ranges[pos][1] + 1, new_time_ranges[new_pos][1] ]
483 new_time_ranges[new_pos][1] = time_ranges[pos][1]
484 new_pos += 1
485 new_time_ranges.insert(new_pos, r)
486 continue
487 # new [start, end] is within old [start, end]
488 new_pos += 1
489
490def SplitTimeRangesByTraceDataDensity(time_ranges, cpus, nr, cmd, file_name, per_cpu, min_size, min_interval, verbosity):
491 if verbosity.normal:
492 print("\rAnalyzing...", flush=True, end=" ")
493 if verbosity.verbose:
494 print()
495 cnts_cmd, times_cmd = PerfDoubleQuickCommands(cmd, file_name)
496
497 nr_cpus = cpus[-1] + 1 if per_cpu else 1
498 if per_cpu:
499 nr_cpus = cpus[-1] + 1
500 cpu_time_ranges = [ CPUTimeRange(cpu) for cpu in range(nr_cpus) ]
501 else:
502 nr_cpus = 1
503 cpu_time_ranges = [ CPUTimeRange(-1) ]
504
505 if verbosity.debug:
506 print("nr_cpus", nr_cpus)
507 print("cnts_cmd", cnts_cmd)
508 print("times_cmd", times_cmd)
509
510 # Count the number of "double quick" samples per CPU
511 ProcessCommandOutputLines(cnts_cmd, per_cpu, CountSamplesByCPU, cpu_time_ranges)
512
513 tot = 0
514 mx = 0
515 for cpu_time_range in cpu_time_ranges:
516 cnt = cpu_time_range.sample_cnt
517 tot += cnt
518 if cnt > mx:
519 mx = cnt
520 if verbosity.debug:
521 print("cpu:", cpu_time_range.cpu, "sample_cnt", cnt)
522
523 if min_size < 1:
524 min_size = 1
525
526 if mx < min_size:
527 # Too little data to be worth splitting
528 if verbosity.debug:
529 print("Too little data to split by time")
530 if nr == 0:
531 nr = 1
532 return [ SplitTimeRangesIntoN(time_ranges, nr, min_interval) ]
533
534 if nr:
535 divisor = nr
536 min_size = 1
537 else:
538 divisor = NumberOfCPUs()
539
540 interval = int(round(tot / divisor, 0))
541 if interval < min_size:
542 interval = min_size
543
544 if verbosity.debug:
545 print("divisor", divisor)
546 print("min_size", min_size)
547 print("interval", interval)
548
549 min_time = time_ranges[0][0]
550 max_time = time_ranges[-1][1]
551
552 for cpu_time_range in cpu_time_ranges:
553 cnt = cpu_time_range.sample_cnt
554 if cnt == 0:
555 cpu_time_range.time_ranges = copy.deepcopy(time_ranges)
556 continue
557 # Adjust target interval for CPU to give approximately equal interval sizes
558 # Determine number of intervals, rounding to nearest integer
559 n = int(round(cnt / interval, 0))
560 if n < 1:
561 n = 1
562 # Determine interval size, rounding up
563 d, m = divmod(cnt, n)
564 if m:
565 d += 1
566 cpu_time_range.interval = d
567 cpu_time_range.interval_remaining = d
568 cpu_time_range.remaining = cnt
569 # Init. time ranges for each CPU with the start time
570 cpu_time_range.time_ranges = [ [min_time, max_time] ]
571
572 # Set time ranges so that the same number of "double quick" samples
573 # will fall into each time range.
574 ProcessCommandOutputLines(times_cmd, per_cpu, CalcTimeRangesByCPU, cpu_time_ranges, max_time)
575
576 for cpu_time_range in cpu_time_ranges:
577 if cpu_time_range.sample_cnt:
578 IntersectTimeRanges(cpu_time_range.time_ranges, time_ranges)
579
580 return [cpu_time_ranges[cpu].time_ranges for cpu in cpus]
581
582def SplitSingleTimeRangeIntoN(time_range, n):
583 if n <= 1:
584 return [time_range]
585 start = time_range[0]
586 end = time_range[1]
587 duration = int((end - start + 1) / n)
588 if duration < 1:
589 return [time_range]
590 time_ranges = []
591 for i in range(n):
592 time_ranges.append([start, start + duration - 1])
593 start += duration
594 time_ranges[-1][1] = end
595 return time_ranges
596
597def TimeRangeDuration(r):
598 return r[1] - r[0] + 1
599
600def TotalDuration(time_ranges):
601 duration = 0
602 for r in time_ranges:
603 duration += TimeRangeDuration(r)
604 return duration
605
606def SplitTimeRangesByInterval(time_ranges, interval):
607 new_ranges = []
608 for r in time_ranges:
609 duration = TimeRangeDuration(r)
610 n = duration / interval
611 n = int(round(n, 0))
612 new_ranges += SplitSingleTimeRangeIntoN(r, n)
613 return new_ranges
614
615def SplitTimeRangesIntoN(time_ranges, n, min_interval):
616 if n <= len(time_ranges):
617 return time_ranges
618 duration = TotalDuration(time_ranges)
619 interval = duration / n
620 if interval < min_interval:
621 interval = min_interval
622 return SplitTimeRangesByInterval(time_ranges, interval)
623
624def RecombineTimeRanges(tr):
625 new_tr = copy.deepcopy(tr)
626 n = len(new_tr)
627 i = 1
628 while i < len(new_tr):
629 # if prev end + 1 == cur start, combine them
630 if new_tr[i - 1][1] + 1 == new_tr[i][0]:
631 new_tr[i][0] = new_tr[i - 1][0]
632 del new_tr[i - 1]
633 else:
634 i += 1
635 return new_tr
636
637def OpenTimeRangeEnds(time_ranges, min_time, max_time):
638 if time_ranges[0][0] <= min_time:
639 time_ranges[0][0] = None
640 if time_ranges[-1][1] >= max_time:
641 time_ranges[-1][1] = None
642
643def BadTimeStr(time_str):
644 raise Exception(f"perf command bad time option: '{time_str}'\nCheck also 'time of first sample' and 'time of last sample' in perf script --header-only")
645
646def ValidateTimeRanges(time_ranges, time_str):
647 n = len(time_ranges)
648 for i in range(n):
649 start = time_ranges[i][0]
650 end = time_ranges[i][1]
651 if i != 0 and start <= time_ranges[i - 1][1]:
652 BadTimeStr(time_str)
653 if start > end:
654 BadTimeStr(time_str)
655
656def TimeVal(s, dflt):
657 s = s.strip()
658 if s == "":
659 return dflt
660 a = s.split(".")
661 if len(a) > 2:
662 raise Exception(f"Bad time value'{s}'")
663 x = int(a[0])
664 if x < 0:
665 raise Exception("Negative time not allowed")
666 x *= 1000000000
667 if len(a) > 1:
668 x += int((a[1] + "000000000")[:9])
669 return x
670
671def BadCPUStr(cpu_str):
672 raise Exception(f"perf command bad cpu option: '{cpu_str}'\nCheck also 'nrcpus avail' in perf script --header-only")
673
674def ParseTimeStr(time_str, min_time, max_time):
675 if time_str == None or time_str == "":
676 return [[min_time, max_time]]
677 time_ranges = []
678 for r in time_str.split():
679 a = r.split(",")
680 if len(a) != 2:
681 BadTimeStr(time_str)
682 try:
683 start = TimeVal(a[0], min_time)
684 end = TimeVal(a[1], max_time)
685 except:
686 BadTimeStr(time_str)
687 time_ranges.append([start, end])
688 ValidateTimeRanges(time_ranges, time_str)
689 return time_ranges
690
691def ParseCPUStr(cpu_str, nr_cpus):
692 if cpu_str == None or cpu_str == "":
693 return [-1]
694 cpus = []
695 for r in cpu_str.split(","):
696 a = r.split("-")
697 if len(a) < 1 or len(a) > 2:
698 BadCPUStr(cpu_str)
699 try:
700 start = int(a[0].strip())
701 if len(a) > 1:
702 end = int(a[1].strip())
703 else:
704 end = start
705 except:
706 BadCPUStr(cpu_str)
707 if start < 0 or end < 0 or end < start or end >= nr_cpus:
708 BadCPUStr(cpu_str)
709 cpus.extend(range(start, end + 1))
710 cpus = list(set(cpus)) # Remove duplicates
711 cpus.sort()
712 return cpus
713
714class ParallelPerf():
715
716 def __init__(self, a):
717 for arg_name in vars(a):
718 setattr(self, arg_name, getattr(a, arg_name))
719 self.orig_nr = self.nr
720 self.orig_cmd = list(self.cmd)
721 self.perf = self.cmd[0]
722 if os.path.exists(self.output_dir):
723 raise Exception(f"Output '{self.output_dir}' already exists")
724 if self.jobs < 0 or self.nr < 0 or self.interval < 0:
725 raise Exception("Bad options (negative values): try -h option for help")
726 if self.nr != 0 and self.interval != 0:
727 raise Exception("Cannot specify number of time subdivisions and time interval")
728 if self.jobs == 0:
729 self.jobs = NumberOfCPUs()
730 if self.nr == 0 and self.interval == 0:
731 if self.per_cpu:
732 self.nr = 1
733 else:
734 self.nr = self.jobs
735
736 def Init(self):
737 if self.verbosity.debug:
738 print("cmd", self.cmd)
739 self.file_name = DetermineInputFileName(self.cmd)
740 self.hdr = ReadHeader(self.perf, self.file_name)
741 self.hdr_dict = ParseHeader(self.hdr)
742 self.cmd_line = HeaderField(self.hdr_dict, "cmdline")
743
744 def ExtractTimeInfo(self):
745 self.min_time = TimeVal(HeaderField(self.hdr_dict, "time of first sample"), 0)
746 self.max_time = TimeVal(HeaderField(self.hdr_dict, "time of last sample"), 0)
747 self.time_str = ExtractPerfOption(self.cmd, "", "time")
748 self.time_ranges = ParseTimeStr(self.time_str, self.min_time, self.max_time)
749 if self.verbosity.debug:
750 print("time_ranges", self.time_ranges)
751
752 def ExtractCPUInfo(self):
753 if self.per_cpu:
754 nr_cpus = int(HeaderField(self.hdr_dict, "nrcpus avail"))
755 self.cpu_str = ExtractPerfOption(self.cmd, "C", "cpu")
756 if self.cpu_str == None or self.cpu_str == "":
757 self.cpus = [ x for x in range(nr_cpus) ]
758 else:
759 self.cpus = ParseCPUStr(self.cpu_str, nr_cpus)
760 else:
761 self.cpu_str = None
762 self.cpus = [-1]
763 if self.verbosity.debug:
764 print("cpus", self.cpus)
765
766 def IsIntelPT(self):
767 return self.cmd_line.find("intel_pt") >= 0
768
769 def SplitTimeRanges(self):
770 if self.IsIntelPT() and self.interval == 0:
771 self.split_time_ranges_for_each_cpu = \
772 SplitTimeRangesByTraceDataDensity(self.time_ranges, self.cpus, self.orig_nr,
773 self.orig_cmd, self.file_name, self.per_cpu,
774 self.min_size, self.min_interval, self.verbosity)
775 elif self.nr:
776 self.split_time_ranges_for_each_cpu = [ SplitTimeRangesIntoN(self.time_ranges, self.nr, self.min_interval) ]
777 else:
778 self.split_time_ranges_for_each_cpu = [ SplitTimeRangesByInterval(self.time_ranges, self.interval) ]
779
780 def CheckTimeRanges(self):
781 for tr in self.split_time_ranges_for_each_cpu:
782 # Re-combined time ranges should be the same
783 new_tr = RecombineTimeRanges(tr)
784 if new_tr != self.time_ranges:
785 if self.verbosity.debug:
786 print("tr", tr)
787 print("new_tr", new_tr)
788 raise Exception("Self test failed!")
789
790 def OpenTimeRangeEnds(self):
791 for time_ranges in self.split_time_ranges_for_each_cpu:
792 OpenTimeRangeEnds(time_ranges, self.min_time, self.max_time)
793
794 def CreateWorkList(self):
795 self.worklist = CreateWorkList(self.cmd, self.pipe_to, self.output_dir, self.cpus, self.split_time_ranges_for_each_cpu)
796
797 def PerfDataRecordedPerCPU(self):
798 if "--per-thread" in self.cmd_line.split():
799 return False
800 return True
801
802 def DefaultToPerCPU(self):
803 # --no-per-cpu option takes precedence
804 if self.no_per_cpu:
805 return False
806 if not self.PerfDataRecordedPerCPU():
807 return False
808 # Default to per-cpu for Intel PT data that was recorded per-cpu,
809 # because decoding can be done for each CPU separately.
810 if self.IsIntelPT():
811 return True
812 return False
813
814 def Config(self):
815 self.Init()
816 self.ExtractTimeInfo()
817 if not self.per_cpu:
818 self.per_cpu = self.DefaultToPerCPU()
819 if self.verbosity.debug:
820 print("per_cpu", self.per_cpu)
821 self.ExtractCPUInfo()
822 self.SplitTimeRanges()
823 if self.verbosity.self_test:
824 self.CheckTimeRanges()
825 # Prefer open-ended time range to starting / ending with min_time / max_time resp.
826 self.OpenTimeRangeEnds()
827 self.CreateWorkList()
828
829 def Run(self):
830 if self.dry_run:
831 print(len(self.worklist),"jobs:")
832 for w in self.worklist:
833 print(w.Command())
834 return True
835 result = RunWork(self.worklist, self.jobs, verbosity=self.verbosity)
836 if self.verbosity.verbose:
837 print(glb_prog_name, "done")
838 return result
839
840def RunParallelPerf(a):
841 pp = ParallelPerf(a)
842 pp.Config()
843 return pp.Run()
844
845def Main(args):
846 ap = argparse.ArgumentParser(
847 prog=glb_prog_name, formatter_class = argparse.RawDescriptionHelpFormatter,
848 description =
849"""
850Run a perf script command multiple times in parallel, using perf script options
851--cpu and --time so that each job processes a different chunk of the data.
852""",
853 epilog =
854"""
855Follow the options by '--' and then the perf script command e.g.
856
857 $ perf record -a -- sleep 10
858 $ parallel-perf.py --nr=4 -- perf script --ns
859 All jobs finished successfully
860 $ tree parallel-perf-output/
861 parallel-perf-output/
862 ├── time-range-0
863 │   ├── cmd.txt
864 │   └── out.txt
865 ├── time-range-1
866 │   ├── cmd.txt
867 │   └── out.txt
868 ├── time-range-2
869 │   ├── cmd.txt
870 │   └── out.txt
871 └── time-range-3
872 ├── cmd.txt
873 └── out.txt
874 $ find parallel-perf-output -name cmd.txt | sort | xargs grep -H .
875 parallel-perf-output/time-range-0/cmd.txt:perf script --time=,9466.504461499 --ns
876 parallel-perf-output/time-range-1/cmd.txt:perf script --time=9466.504461500,9469.005396999 --ns
877 parallel-perf-output/time-range-2/cmd.txt:perf script --time=9469.005397000,9471.506332499 --ns
878 parallel-perf-output/time-range-3/cmd.txt:perf script --time=9471.506332500, --ns
879
880Any perf script command can be used, including the use of perf script options
881--dlfilter and --script, so that the benefit of running parallel jobs
882naturally extends to them also.
883
884If option --pipe-to is used, standard output is first piped through that
885command. Beware, if the command fails (e.g. grep with no matches), it will be
886considered a fatal error.
887
888Final standard output is redirected to files named out.txt in separate
889subdirectories under the output directory. Similarly, standard error is
890written to files named err.txt. In addition, files named cmd.txt contain the
891corresponding perf script command. After processing, err.txt files are removed
892if they are empty.
893
894If any job exits with a non-zero exit code, then all jobs are killed and no
895more are started. A message is printed if any job results in a non-empty
896err.txt file.
897
898There is a separate output subdirectory for each time range. If the --per-cpu
899option is used, these are further grouped under cpu-n subdirectories, e.g.
900
901 $ parallel-perf.py --per-cpu --nr=2 -- perf script --ns --cpu=0,1
902 All jobs finished successfully
903 $ tree parallel-perf-output
904 parallel-perf-output/
905 ├── cpu-0
906 │   ├── time-range-0
907 │   │   ├── cmd.txt
908 │   │   └── out.txt
909 │   └── time-range-1
910 │   ├── cmd.txt
911 │   └── out.txt
912 └── cpu-1
913 ├── time-range-0
914 │   ├── cmd.txt
915 │   └── out.txt
916 └── time-range-1
917 ├── cmd.txt
918 └── out.txt
919 $ find parallel-perf-output -name cmd.txt | sort | xargs grep -H .
920 parallel-perf-output/cpu-0/time-range-0/cmd.txt:perf script --cpu=0 --time=,9469.005396999 --ns
921 parallel-perf-output/cpu-0/time-range-1/cmd.txt:perf script --cpu=0 --time=9469.005397000, --ns
922 parallel-perf-output/cpu-1/time-range-0/cmd.txt:perf script --cpu=1 --time=,9469.005396999 --ns
923 parallel-perf-output/cpu-1/time-range-1/cmd.txt:perf script --cpu=1 --time=9469.005397000, --ns
924
925Subdivisions of time range, and cpus if the --per-cpu option is used, are
926expressed by the --time and --cpu perf script options respectively. If the
927supplied perf script command has a --time option, then that time range is
928subdivided, otherwise the time range given by 'time of first sample' to
929'time of last sample' is used (refer perf script --header-only). Similarly, the
930supplied perf script command may provide a --cpu option, and only those CPUs
931will be processed.
932
933To prevent time intervals becoming too small, the --min-interval option can
934be used.
935
936Note there is special handling for processing Intel PT traces. If an interval is
937not specified and the perf record command contained the intel_pt event, then the
938time range will be subdivided in order to produce subdivisions that contain
939approximately the same amount of trace data. That is accomplished by counting
940double-quick (--itrace=qqi) samples, and choosing time ranges that encompass
941approximately the same number of samples. In that case, time ranges may not be
942the same for each CPU processed. For Intel PT, --per-cpu is the default, but
943that can be overridden by --no-per-cpu. Note, for Intel PT, double-quick
944decoding produces 1 sample for each PSB synchronization packet, which in turn
945come after a certain number of bytes output, determined by psb_period (refer
946perf Intel PT documentation). The minimum number of double-quick samples that
947will define a time range can be set by the --min_size option, which defaults to
94864.
949""")
950 ap.add_argument("-o", "--output-dir", default="parallel-perf-output", help="output directory (default 'parallel-perf-output')")
951 ap.add_argument("-j", "--jobs", type=int, default=0, help="maximum number of jobs to run in parallel at one time (default is the number of CPUs)")
952 ap.add_argument("-n", "--nr", type=int, default=0, help="number of time subdivisions (default is the number of jobs)")
953 ap.add_argument("-i", "--interval", type=float, default=0, help="subdivide the time range using this time interval (in seconds e.g. 0.1 for a tenth of a second)")
954 ap.add_argument("-c", "--per-cpu", action="store_true", help="process data for each CPU in parallel")
955 ap.add_argument("-m", "--min-interval", type=float, default=glb_min_interval, help=f"minimum interval (default {glb_min_interval} seconds)")
956 ap.add_argument("-p", "--pipe-to", help="command to pipe output to (optional)")
957 ap.add_argument("-N", "--no-per-cpu", action="store_true", help="do not process data for each CPU in parallel")
958 ap.add_argument("-b", "--min_size", type=int, default=glb_min_samples, help="minimum data size (for Intel PT in PSBs)")
959 ap.add_argument("-D", "--dry-run", action="store_true", help="do not run any jobs, just show the perf script commands")
960 ap.add_argument("-q", "--quiet", action="store_true", help="do not print any messages except errors")
961 ap.add_argument("-v", "--verbose", action="store_true", help="print more messages")
962 ap.add_argument("-d", "--debug", action="store_true", help="print debugging messages")
963 cmd_line = list(args)
964 try:
965 split_pos = cmd_line.index("--")
966 cmd = cmd_line[split_pos + 1:]
967 args = cmd_line[:split_pos]
968 except:
969 cmd = None
970 args = cmd_line
971 a = ap.parse_args(args=args[1:])
972 a.cmd = cmd
973 a.verbosity = Verbosity(a.quiet, a.verbose, a.debug)
974 try:
975 if a.cmd == None:
976 if len(args) <= 1:
977 ap.print_help()
978 return True
979 raise Exception("Command line must contain '--' before perf command")
980 return RunParallelPerf(a)
981 except Exception as e:
982 print("Fatal error: ", str(e))
983 if a.debug:
984 raise
985 return False
986
987if __name__ == "__main__":
988 if not Main(sys.argv):
989 sys.exit(1)