Add a Python script to run a perf script command multiple times in
parallel, using perf script options --cpu and --time so that each job
processes a different chunk of the data.
Refer to the script's own help text at the end of the patch for more
details.
The script is useful for Intel PT traces, that can be efficiently
decoded by perf script when split by CPU and/or time ranges. Running
jobs in parallel can decrease the overall decoding time.
Signed-off-by: Adrian Hunter <[email protected]>
---
tools/perf/scripts/python/parallel-perf.py | 976 +++++++++++++++++++++
1 file changed, 976 insertions(+)
create mode 100755 tools/perf/scripts/python/parallel-perf.py
diff --git a/tools/perf/scripts/python/parallel-perf.py b/tools/perf/scripts/python/parallel-perf.py
new file mode 100755
index 000000000000..9a2939c5afb5
--- /dev/null
+++ b/tools/perf/scripts/python/parallel-perf.py
@@ -0,0 +1,976 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: GPL-2.0
+#
+# Run a perf script command multiple times in parallel, using perf script
+# options --cpu and --time so that each job processes a different chunk
+# of the data.
+#
+# Copyright (c) 2024, Intel Corporation.
+
+import subprocess
+import argparse
+import pathlib
+import shlex
+import time
+import copy
+import sys
+import os
+import re
+
+glb_prog_name = "parallel-perf.py"
+glb_min_samples = 64
+
+class Verbosity():
+
+ def __init__(self):
+ self.normal = True
+ self.verbose = False
+ self.debug = False
+ self.self_test = True
+
+ def Quiet(self):
+ self.normal = False
+ self.verbose = False
+ self.debug = False
+
+ def Verbose(self):
+ self.normal = True
+ self.verbose = True
+ self.debug = False
+
+ def Debug(self):
+ self.normal = True
+ self.verbose = True
+ self.debug = True
+
+ def VerboseNewLine(self):
+ if self.verbose:
+ print()
+
+ def NormalNewLine(self):
+ if self.verbose:
+ print()
+
+class Work():
+
+ def __init__(self, cmd, output_dir="."):
+ self.popen = None
+ self.cmd = cmd
+ self.output_dir = output_dir
+ self.cmdout_name = output_dir + "/cmd.txt"
+ self.stdout_name = output_dir + "/out.txt"
+ self.stderr_name = output_dir + "/err.txt"
+
+ def Command(self):
+ sh_cmd = [ shlex.quote(x) for x in self.cmd ]
+ return " ".join(self.cmd)
+
+ def Stdout(self):
+ return open(self.stdout_name, "w")
+
+ def Stderr(self):
+ return open(self.stderr_name, "w")
+
+ def CreateOutputDir(self):
+ pathlib.Path(self.output_dir).mkdir(parents=True, exist_ok=True)
+
+ def Start(self):
+ if self.popen:
+ return
+ self.CreateOutputDir()
+ with open(self.cmdout_name, "w") as f:
+ f.write(self.Command())
+ f.write("\n")
+ stdout = self.Stdout()
+ stderr = self.Stderr()
+ self.popen = subprocess.Popen(self.cmd, stdout=stdout, stderr=stderr)
+
+ def RemoveEmptyErrFile(self):
+ if os.path.exists(self.stderr_name):
+ if os.path.getsize(self.stderr_name) == 0:
+ os.unlink(self.stderr_name)
+
+ def Errors(self):
+ if os.path.exists(self.stderr_name):
+ if os.path.getsize(self.stderr_name) != 0:
+ return [ "Non-empty error file " + self.stderr_name ]
+ return []
+
+ def TidyUp(self):
+ self.RemoveEmptyErrFile()
+
+ def Poll(self, wait=False):
+ if not self.popen:
+ return None
+ if self.popen.returncode != None:
+ return self.popen.returncode
+ if wait:
+ result = self.popen.wait()
+ else:
+ result = self.popen.poll()
+ if result != None:
+ self.TidyUp()
+ return result
+
+ def Wait(self):
+ return self.Poll(wait=True)
+
+ def Kill(self):
+ if not self.popen:
+ return
+ self.popen.kill()
+
+def KillWork(worklist, verbosity):
+ for w in worklist:
+ w.Kill()
+ for w in worklist:
+ w.Wait()
+
+def NumberOfCPUs():
+ return os.sysconf("SC_NPROCESSORS_ONLN")
+
+def NanoSecsToSecsStr(x):
+ if x == None:
+ return ""
+ x = str(x)
+ if len(x) < 10:
+ x = "0" * (10 - len(x)) + x
+ return x[:len(x) - 9] + "." + x[-9:]
+
+def InsertOptionAfter(cmd, option, after):
+ try:
+ pos = cmd.index(after)
+ cmd.insert(pos + 1, option)
+ except:
+ cmd.append(option)
+
+def CreateWorkList(cmd, output_dir, cpus, time_ranges_by_cpu):
+ max_len = len(str(cpus[-1]))
+ cpu_dir_fmt = "cpu-%." + str(max_len) + "u"
+ worklist = []
+ pos = 0
+ for cpu in cpus:
+ if cpu >= 0:
+ cpu_dir = os.path.join(output_dir, cpu_dir_fmt % cpu)
+ cpu_option = "--cpu=" + str(cpu)
+ else:
+ cpu_dir = output_dir
+ cpu_option = None
+
+ tr_dir_fmt = "time-range"
+
+ if len(time_ranges_by_cpu) > 1:
+ time_ranges = time_ranges_by_cpu[pos]
+ tr_dir_fmt += "-" + str(pos)
+ pos += 1
+ else:
+ time_ranges = time_ranges_by_cpu[0]
+
+ max_len = len(str(len(time_ranges)))
+ tr_dir_fmt += "-%." + str(max_len) + "u"
+
+ i = 0
+ for r in time_ranges:
+ if r == [None, None]:
+ time_option = None
+ work_output_dir = cpu_dir
+ else:
+ time_option = "--time=" + NanoSecsToSecsStr(r[0]) + "," + NanoSecsToSecsStr(r[1])
+ work_output_dir = os.path.join(cpu_dir, tr_dir_fmt % i)
+ i += 1
+ work_cmd = list(cmd)
+ if time_option != None:
+ InsertOptionAfter(work_cmd, time_option, "script")
+ if cpu_option != None:
+ InsertOptionAfter(work_cmd, cpu_option, "script")
+ w = Work(work_cmd, work_output_dir)
+ worklist.append(w)
+ return worklist
+
+def DoRunWork(worklist, nr_jobs, verbosity):
+ nr_to_do = len(worklist)
+ not_started = list(worklist)
+ running = []
+ done = []
+ chg = False
+ while True:
+ nr_done = len(done)
+ if chg and verbosity.normal:
+ nr_run = len(running)
+ print(f"\rThere are {nr_to_do} jobs: {nr_done} completed, {nr_run} running", flush=True, end=" ")
+ verbosity.VerboseNewLine()
+ chg = False
+ if nr_done == nr_to_do:
+ break
+ while len(running) < nr_jobs and len(not_started):
+ w = not_started.pop(0)
+ running.append(w)
+ if verbosity.verbose:
+ print("Starting:", w.Command())
+ w.Start()
+ chg = True
+ if len(running):
+ time.sleep(0.1)
+ finished = []
+ not_finished = []
+ while len(running):
+ w = running.pop(0)
+ r = w.Poll()
+ if r == None:
+ not_finished.append(w)
+ continue
+ if r == 0:
+ if verbosity.verbose:
+ print("Finished:", w.Command())
+ finished.append(w)
+ chg = True
+ continue
+ verbosity.NormalNewLine()
+ print("Job failed! return code:", r, "command:", w.Command())
+ print("Killing outstanding jobs")
+ KillWork(not_finished, verbosity)
+ KillWork(running, verbosity)
+ return False
+ running = not_finished
+ done += finished
+ errorlist = []
+ for w in worklist:
+ errorlist += w.Errors()
+ if len(errorlist):
+ print("Errors:")
+ for e in errorlist:
+ print(e)
+ elif verbosity.normal:
+ print("\r"," "*50, "\rAll jobs finished successfully", flush=True)
+ return True
+
+def RunWork(worklist, nr_jobs=NumberOfCPUs(), verbosity=Verbosity()):
+ try:
+ return DoRunWork(worklist, nr_jobs, verbosity)
+ except:
+ for w in worklist:
+ w.Kill()
+ raise
+ return True
+
+def ReadHeader(perf, file_name):
+ return subprocess.Popen([perf, "script", "--header-only", "--input", file_name], stdout=subprocess.PIPE).stdout.read().decode("utf-8")
+
+def ParseHeader(hdr):
+ result = {}
+ lines = hdr.split("\n")
+ for line in lines:
+ if ":" in line and line[0] == "#":
+ pos = line.index(":")
+ name = line[1:pos-1].strip()
+ value = line[pos+1:].strip()
+ if name in result:
+ orig_name = name
+ nr = 2
+ while True:
+ name = orig_name + " " + str(nr)
+ if name not in result:
+ break
+ nr += 1
+ result[name] = value
+ return result
+
+def HeaderField(hdr_dict, hdr_fld):
+ if hdr_fld not in hdr_dict:
+ raise Exception("'" + hdr_fld + "' missing from header information")
+ return hdr_dict[hdr_fld]
+
+class OptPos():
+
+ def Init(self, opt_element=-1, value_element=-1, opt_pos=-1, value_pos=-1, error=None):
+ self.opt_element = opt_element # list element that contains option
+ self.value_element = value_element # list element that contains option value
+ self.opt_pos = opt_pos # string position of option
+ self.value_pos = value_pos # string position of value
+ self.error = error # error message string
+
+ def __init__(self, args, short_name, long_name, default=None):
+ self.args = list(args)
+ self.default = default
+ n = 2 + len(long_name)
+ m = len(short_name)
+ pos = -1
+ for opt in args:
+ pos += 1
+ if m and opt[:2] == "-" + short_name:
+ if len(opt) == 2:
+ if pos + 1 < len(args):
+ self.Init(pos, pos + 1, 0, 0)
+ else:
+ self.Init(error = "-" + short_name + " option missing value")
+ else:
+ self.Init(pos, pos, 0, 2)
+ return
+ if opt[:n] == "--" + long_name:
+ if len(opt) == n:
+ if pos + 1 < len(args):
+ self.Init(pos, pos + 1, 0, 0)
+ else:
+ self.Init(error = "--" + long_name + " option missing value")
+ elif opt[n] == "=":
+ self.Init(pos, pos, 0, n + 1)
+ else:
+ self.Init(error = "--" + long_name + " option expected '='")
+ return
+ if m and opt[:1] == "-" and opt[:2] != "--" and short_name in opt:
+ ipos = opt.index(short_name)
+ if "-" in opt[1:]:
+ hpos = opt[1:].index("-")
+ if hpos < ipos:
+ continue
+ if ipos + 1 == len(opt):
+ if pos + 1 < len(args):
+ self.Init(pos, pos + 1, ipos, 0)
+ else:
+ self.Init(error = "-" + short_name + " option missing value")
+ else:
+ self.Init(pos, pos, ipos, ipos + 1)
+ return
+ self.Init()
+
+ def Value(self):
+ if self.opt_element >= 0:
+ if self.opt_element != self.value_element:
+ return self.args[self.value_element]
+ else:
+ return self.args[self.value_element][self.value_pos:]
+ return self.default
+
+ def Remove(self, args):
+ if self.opt_element == -1:
+ return
+ if self.opt_element != self.value_element:
+ del args[self.value_element]
+ if self.opt_pos:
+ args[self.opt_element] = args[self.opt_element][:self.opt_pos]
+ else:
+ del args[self.opt_element]
+
+def DetermineInputFileName(cmd):
+ p = OptPos(cmd, "i", "input", "perf.data")
+ if p.error:
+ raise Exception("perf command " + p.error)
+ file_name = p.Value()
+ if not os.path.exists(file_name):
+ raise Exception("perf command input file '" + file_name + "' not found")
+ return file_name
+
+def ReadOption(args, short_name, long_name, err_prefix, remove=False):
+ p = OptPos(args, short_name, long_name)
+ if p.error:
+ raise Exception(err_prefix + p.error)
+ value = p.Value()
+ if remove:
+ p.Remove(args)
+ return value
+
+def ExtractOption(args, short_name, long_name, err_prefix):
+ return ReadOption(args, short_name, long_name, err_prefix, True)
+
+def ReadPerfOption(args, short_name, long_name):
+ return ReadOption(args, short_name, long_name, "perf command ")
+
+def ExtractPerfOption(args, short_name, long_name):
+ return ExtractOption(args, short_name, long_name, "perf command ")
+
+def PerfDoubleQuickCommands(cmd, file_name):
+ cpu_str = ReadPerfOption(cmd, "C", "cpu")
+ time_str = ReadPerfOption(cmd, "", "time")
+ # Use double-quick sampling to determine trace data density
+ times_cmd = ["perf", "script", "--ns", "--input", file_name, "--itrace=qqi"]
+ if cpu_str != None and cpu_str != "":
+ times_cmd.append("--cpu=" + cpu_str)
+ if time_str != None and time_str != "":
+ times_cmd.append("--time=" + time_str)
+ cnts_cmd = list(times_cmd)
+ cnts_cmd.append("-Fcpu")
+ times_cmd.append("-Fcpu,time")
+ return cnts_cmd, times_cmd
+
+class CPUTimeRange():
+ def __init__(self, cpu):
+ self.cpu = cpu
+ self.sample_cnt = 0
+ self.time_ranges = None
+ self.interval = 0
+ self.interval_remaining = 0
+ self.remaining = 0
+ self.tr_pos = 0
+
+def CalcTimeRangesByCPU(line, cpu, cpu_time_ranges, max_time):
+ cpu_time_range = cpu_time_ranges[cpu]
+ cpu_time_range.remaining -= 1
+ cpu_time_range.interval_remaining -= 1
+ if cpu_time_range.remaining == 0:
+ cpu_time_range.time_ranges[cpu_time_range.tr_pos][1] = max_time
+ return
+ if cpu_time_range.interval_remaining == 0:
+ time = TimeVal(line[1][:-1], 0)
+ time_ranges = cpu_time_range.time_ranges
+ time_ranges[cpu_time_range.tr_pos][1] = time - 1
+ time_ranges.append([time, max_time])
+ cpu_time_range.tr_pos += 1
+ cpu_time_range.interval_remaining = cpu_time_range.interval
+
+def CountSamplesByCPU(line, cpu, cpu_time_ranges):
+ try:
+ cpu_time_ranges[cpu].sample_cnt += 1
+ except:
+ print("exception")
+ print("cpu", cpu)
+ print("len(cpu_time_ranges)", len(cpu_time_ranges))
+ raise
+
+def ProcessCommandOutputLines(cmd, per_cpu, fn, *x):
+ # Assume CPU number is at beginning of line and enclosed by []
+ pat = re.compile(r"\s*\[[0-9]+\]")
+ p = subprocess.Popen(cmd, stdout=subprocess.PIPE)
+ while True:
+ if line := p.stdout.readline():
+ line = line.decode("utf-8")
+ if pat.match(line):
+ line = line.split()
+ if per_cpu:
+ # Assumes CPU number is enclosed by []
+ cpu = int(line[0][1:-1])
+ else:
+ cpu = 0
+ fn(line, cpu, *x)
+ else:
+ break
+ p.wait()
+
+def IntersectTimeRanges(new_time_ranges, time_ranges):
+ pos = 0
+ new_pos = 0
+ # Can assume len(time_ranges) != 0 and len(new_time_ranges) != 0
+ # Note also, there *must* be at least one intersection.
+ while pos < len(time_ranges) and new_pos < len(new_time_ranges):
+ # new end < old start => no intersection, remove new
+ if new_time_ranges[new_pos][1] < time_ranges[pos][0]:
+ del new_time_ranges[new_pos]
+ continue
+ # new start > old end => no intersection, check next
+ if new_time_ranges[new_pos][0] > time_ranges[pos][1]:
+ pos += 1
+ if pos < len(time_ranges):
+ continue
+ # no next, so remove remaining
+ while new_pos < len(new_time_ranges):
+ del new_time_ranges[new_pos]
+ return
+ # Found an intersection
+ # new start < old start => adjust new start = old start
+ if new_time_ranges[new_pos][0] < time_ranges[pos][0]:
+ new_time_ranges[new_pos][0] = time_ranges[pos][0]
+ # new end > old end => keep the overlap, insert the remainder
+ if new_time_ranges[new_pos][1] > time_ranges[pos][1]:
+ r = [ time_ranges[pos][1] + 1, new_time_ranges[new_pos][1] ]
+ new_time_ranges[new_pos][1] = time_ranges[pos][1]
+ new_pos += 1
+ new_time_ranges.insert(new_pos, r)
+ continue
+ # new [start, end] is within old [start, end]
+ new_pos += 1
+
+def SplitTimeRangesByTraceDataDensity(time_ranges, cpus, nr, cmd, file_name, per_cpu, minimum, verbosity):
+ if verbosity.normal:
+ print("\rAnalyzing...", flush=True, end=" ")
+ verbosity.VerboseNewLine()
+ cnts_cmd, times_cmd = PerfDoubleQuickCommands(cmd, file_name)
+
+ nr_cpus = cpus[-1] + 1 if per_cpu else 1
+ if per_cpu:
+ nr_cpus = cpus[-1] + 1
+ cpu_time_ranges = [ CPUTimeRange(cpu) for cpu in range(nr_cpus) ]
+ else:
+ nr_cpus = 1
+ cpu_time_ranges = [ CPUTimeRange(-1) ]
+
+ if verbosity.debug:
+ print("nr_cpus", nr_cpus)
+ print("cnts_cmd", cnts_cmd)
+ print("times_cmd", times_cmd)
+
+ # Count the number of "double quick" samples per CPU
+ ProcessCommandOutputLines(cnts_cmd, per_cpu, CountSamplesByCPU, cpu_time_ranges)
+
+ tot = 0
+ mx = 0
+ for cpu_time_range in cpu_time_ranges:
+ cnt = cpu_time_range.sample_cnt
+ tot += cnt
+ if cnt > mx:
+ mx = cnt
+ if verbosity.debug:
+ print("cpu:", cpu_time_range.cpu, "sample_cnt", cnt)
+
+ if minimum < 1:
+ minimum = 1
+
+ if mx < minimum:
+ # Too little data to be worth splitting
+ if verbosity.debug:
+ print("Too little data to split by time")
+ if nr == 0:
+ nr = 1
+ return [ SplitTimeRangesIntoN(time_ranges, nr) ]
+
+ if nr:
+ divisor = nr
+ minimum = 1
+ else:
+ divisor = NumberOfCPUs()
+
+ interval = int(round(tot / divisor, 0))
+ if interval < minimum:
+ interval = minimum
+
+ if verbosity.debug:
+ print("divisor", divisor)
+ print("minimum", minimum)
+ print("interval", interval)
+
+ min_time = time_ranges[0][0]
+ max_time = time_ranges[-1][1]
+
+ for cpu_time_range in cpu_time_ranges:
+ cnt = cpu_time_range.sample_cnt
+ if cnt == 0:
+ cpu_time_range.time_ranges = copy.deepcopy(time_ranges)
+ continue
+ # Adjust target interval for CPU to give approximately equal interval sizes
+ # Determine number of intervals, rounding to nearest integer
+ n = int(round(cnt / interval, 0))
+ if n < 1:
+ n = 1
+ # Determine interval size, rounding up
+ d, m = divmod(cnt, n)
+ if m:
+ d += 1
+ cpu_time_range.interval = d
+ cpu_time_range.interval_remaining = d
+ cpu_time_range.remaining = cnt
+ # Init. time ranges for each CPU with the start time
+ cpu_time_range.time_ranges = [ [min_time, max_time] ]
+
+ # Set time ranges so that the same number of "double quick" samples
+ # will fall into each time range.
+ ProcessCommandOutputLines(times_cmd, per_cpu, CalcTimeRangesByCPU, cpu_time_ranges, max_time)
+
+ for cpu_time_range in cpu_time_ranges:
+ if cpu_time_range.sample_cnt:
+ IntersectTimeRanges(cpu_time_range.time_ranges, time_ranges)
+
+ return [cpu_time_ranges[cpu].time_ranges for cpu in cpus]
+
+def SplitSingleTimeRangeIntoN(time_range, n):
+ if n <= 1:
+ return [time_range]
+ start = time_range[0]
+ end = time_range[1]
+ duration = int((end - start + 1) / n)
+ if duration < 1:
+ return [time_range]
+ time_ranges = []
+ for i in range(n):
+ time_ranges.append([start, start + duration - 1])
+ start += duration
+ time_ranges[-1][1] = end
+ return time_ranges
+
+def TimeRangeDuration(r):
+ return r[1] - r[0] + 1
+
+def TotalDuration(time_ranges):
+ duration = 0
+ for r in time_ranges:
+ duration += TimeRangeDuration(r)
+ return duration
+
+def SplitTimeRangesByInterval(time_ranges, interval):
+ new_ranges = []
+ for r in time_ranges:
+ duration = TimeRangeDuration(r)
+ n = duration / interval
+ n = int(round(n, 0))
+ new_ranges += SplitSingleTimeRangeIntoN(r, n)
+ return new_ranges
+
+def SplitTimeRangesIntoN(time_ranges, n):
+ if n <= len(time_ranges):
+ return time_ranges
+ duration = TotalDuration(time_ranges)
+ interval = duration / n
+ return SplitTimeRangesByInterval(time_ranges, interval)
+
+def RecombineTimeRanges(tr):
+ new_tr = copy.deepcopy(tr)
+ n = len(new_tr)
+ i = 1
+ while i < len(new_tr):
+ # if prev end + 1 == cur start, combine them
+ if new_tr[i - 1][1] + 1 == new_tr[i][0]:
+ new_tr[i][0] = new_tr[i - 1][0]
+ del new_tr[i - 1]
+ else:
+ i += 1
+ return new_tr
+
+def OpenTimeRangeEnds(time_ranges, min_time, max_time):
+ if time_ranges[0][0] <= min_time:
+ time_ranges[0][0] = None
+ if time_ranges[-1][1] >= max_time:
+ time_ranges[-1][1] = None
+
+def BadTimeStr(time_str):
+ raise Exception("perf command bad time option: '" + time_str + "'\nCheck also 'time of first sample' and 'time of last sample' in perf script --header-only")
+
+def ValidateTimeRanges(time_ranges, time_str):
+ n = len(time_ranges)
+ for i in range(n):
+ start = time_ranges[i][0]
+ end = time_ranges[i][1]
+ if i != 0 and start <= time_ranges[i - 1][1]:
+ BadTimeStr(time_str)
+ if start > end:
+ BadTimeStr(time_str)
+
+def TimeVal(s, dflt):
+ s = s.strip()
+ if s == "":
+ return dflt
+ a = s.split(".")
+ if len(a) > 2:
+ raise Exception("Bad time value'" + s + "'")
+ x = int(a[0])
+ if x < 0:
+ raise Exception("Negative time not allowed")
+ x *= 1000000000
+ if len(a) > 1:
+ x += int((a[1] + "000000000")[:9])
+ return x
+
+def BadCPUStr(cpu_str):
+ raise Exception("perf command bad cpu option: '" + cpu_str + "'\nCheck also 'nrcpus avail' in perf script --header-only")
+
+def ParseTimeStr(time_str, min_time, max_time):
+ if time_str == None or time_str == "":
+ return [[min_time, max_time]]
+ time_ranges = []
+ for r in time_str.split():
+ a = r.split(",")
+ if len(a) != 2:
+ BadTimeStr(time_str)
+ try:
+ start = TimeVal(a[0], min_time)
+ end = TimeVal(a[1], max_time)
+ except:
+ BadTimeStr(time_str)
+ time_ranges.append([start, end])
+ ValidateTimeRanges(time_ranges, time_str)
+ return time_ranges
+
+def ParseCPUStr(cpu_str, nr_cpus):
+ if cpu_str == None or cpu_str == "":
+ return [-1]
+ cpus = []
+ for r in cpu_str.split(","):
+ a = r.split("-")
+ if len(a) < 1 or len(a) > 2:
+ BadCPUStr(cpu_str)
+ try:
+ start = int(a[0].strip())
+ if len(a) > 1:
+ end = int(a[1].strip())
+ else:
+ end = start
+ except:
+ BadCPUStr(cpu_str)
+ if start < 0 or end < 0 or end < start or end >= nr_cpus:
+ BadCPUStr(cpu_str)
+ cpus.extend(range(start, end + 1))
+ cpus = list(set(cpus)) # Remove duplicates
+ cpus.sort()
+ return cpus
+
+class ParallelPerf():
+
+ def __init__(self, output_dir, cmd, jobs=0, nr=0, interval=0, per_cpu=False, no_per_cpu=False, minimum=glb_min_samples, dry_run=False, verbosity=Verbosity()):
+ if os.path.exists(output_dir):
+ raise Exception("Output '" + output_dir + "' already exists")
+ if jobs < 0 or nr < 0 or interval < 0:
+ raise Exception("Bad options (negative values): try -h option for help")
+ if nr != 0 and interval != 0:
+ raise Exception("Cannot specify number of time subdivisions and time interval")
+ if jobs == 0:
+ jobs = NumberOfCPUs()
+ orig_nr = nr
+ if nr == 0 and interval == 0:
+ if per_cpu:
+ nr = 1
+ else:
+ nr = jobs
+ self.output_dir = output_dir
+ self.orig_cmd = list(cmd)
+ self.cmd = cmd
+ self.perf = cmd[0]
+ self.jobs = jobs
+ self.orig_nr = orig_nr
+ self.nr = nr
+ self.interval = interval
+ self.per_cpu = per_cpu
+ self.no_per_cpu = no_per_cpu
+ self.minimum = minimum
+ self.dry_run = dry_run
+ self.verbosity = verbosity
+
+ def Init(self):
+ if self.verbosity.debug:
+ print("cmd", self.cmd)
+ self.file_name = DetermineInputFileName(self.cmd)
+ self.hdr = ReadHeader(self.perf, self.file_name)
+ self.hdr_dict = ParseHeader(self.hdr)
+ self.cmd_line = HeaderField(self.hdr_dict, "cmdline")
+
+ def ExtractTimeInfo(self):
+ self.min_time = TimeVal(HeaderField(self.hdr_dict, "time of first sample"), 0)
+ self.max_time = TimeVal(HeaderField(self.hdr_dict, "time of last sample"), 0)
+ self.time_str = ExtractPerfOption(self.cmd, "", "time")
+ self.time_ranges = ParseTimeStr(self.time_str, self.min_time, self.max_time)
+ if self.verbosity.debug:
+ print("time_ranges", self.time_ranges)
+
+ def ExtractCPUInfo(self):
+ if self.per_cpu:
+ nr_cpus = int(HeaderField(self.hdr_dict, "nrcpus avail"))
+ self.cpu_str = ExtractPerfOption(self.cmd, "C", "cpu")
+ if self.cpu_str == None or self.cpu_str == "":
+ self.cpus = [ x for x in range(nr_cpus) ]
+ else:
+ self.cpus = ParseCPUStr(self.cpu_str, nr_cpus)
+ else:
+ self.cpu_str = None
+ self.cpus = [-1]
+ if self.verbosity.debug:
+ print("cpus", self.cpus)
+
+ def IsIntelPT(self):
+ return self.cmd_line.find("intel_pt") >= 0
+
+ def SplitTimeRanges(self):
+ if self.IsIntelPT() and self.interval == 0:
+ self.split_time_ranges_for_each_cpu = \
+ SplitTimeRangesByTraceDataDensity(self.time_ranges, self.cpus, self.orig_nr,
+ self.orig_cmd, self.file_name, self.per_cpu,
+ self.minimum, self.verbosity)
+ elif self.nr:
+ self.split_time_ranges_for_each_cpu = [ SplitTimeRangesIntoN(self.time_ranges, self.nr) ]
+ else:
+ self.split_time_ranges_for_each_cpu = [ SplitTimeRangesByInterval(self.time_ranges, self.interval) ]
+
+ def CheckTimeRanges(self):
+ for tr in self.split_time_ranges_for_each_cpu:
+ # Re-combined time ranges should be the same
+ new_tr = RecombineTimeRanges(tr)
+ if new_tr != self.time_ranges:
+ if self.verbosity.debug:
+ print("tr", tr)
+ print("new_tr", new_tr)
+ raise Exception("Self test failed!")
+
+ def OpenTimeRangeEnds(self):
+ for time_ranges in self.split_time_ranges_for_each_cpu:
+ OpenTimeRangeEnds(time_ranges, self.min_time, self.max_time)
+
+ def CreateWorkList(self):
+ self.worklist = CreateWorkList(self.cmd, self.output_dir, self.cpus, self.split_time_ranges_for_each_cpu)
+
+ def PerfDataRecordedPerCPU(self):
+ if "--per-thread" in self.cmd_line.split():
+ return False
+ return True
+
+ def DefaultToPerCPU(self):
+ # --no-per-cpu option takes precedence
+ if self.no_per_cpu:
+ return False
+ if not self.PerfDataRecordedPerCPU():
+ return False
+ # Default to per-cpu for Intel PT data that was recorded per-cpu,
+ # because decoding can be done for each CPU separately.
+ if self.IsIntelPT():
+ return True
+ return False
+
+ def Config(self):
+ self.Init()
+ self.ExtractTimeInfo()
+ if not self.per_cpu:
+ self.per_cpu = self.DefaultToPerCPU()
+ if self.verbosity.debug:
+ print("per_cpu", self.per_cpu)
+ self.ExtractCPUInfo()
+ self.SplitTimeRanges()
+ if self.verbosity.self_test:
+ self.CheckTimeRanges()
+ # Prefer open-ended time range to starting / ending with min_time / max_time resp.
+ self.OpenTimeRangeEnds()
+ self.CreateWorkList()
+
+ def Run(self):
+ if self.dry_run:
+ print(len(self.worklist),"jobs:")
+ for w in self.worklist:
+ print(w.Command())
+ return
+ RunWork(self.worklist, verbosity=self.verbosity)
+ if self.verbosity.verbose:
+ print(glb_prog_name, "done")
+
+def RunParallelPerf(output_dir, cmd, jobs=0, nr=0, interval=0, per_cpu=False, no_per_cpu=False, minimum=glb_min_samples, dry_run=False, verbosity=Verbosity()):
+ pp = ParallelPerf(output_dir, cmd, jobs, nr, interval, per_cpu, no_per_cpu, minimum, dry_run, verbosity)
+ pp.Config()
+ pp.Run()
+
+def Main(args):
+ ap = argparse.ArgumentParser(
+ prog=glb_prog_name, formatter_class = argparse.RawDescriptionHelpFormatter,
+ description =
+"""
+Run a perf script command multiple times in parallel, using perf script options
+--cpu and --time so that each job processes a different chunk of the data.
+""",
+ epilog =
+"""
+Follow the options by '--' and then the perf script command e.g.
+
+ $ perf record -a -- sleep 10
+ $ parallel-perf.py --nr=4 -- perf script --ns
+ All jobs finished successfully
+ $ tree parallel-perf-output/
+ parallel-perf-output/
+ ├── time-range-0
+ │ ├── cmd.txt
+ │ └── out.txt
+ ├── time-range-1
+ │ ├── cmd.txt
+ │ └── out.txt
+ ├── time-range-2
+ │ ├── cmd.txt
+ │ └── out.txt
+ └── time-range-3
+ ├── cmd.txt
+ └── out.txt
+ $ find parallel-perf-output -name cmd.txt | sort | xargs grep -H .
+ parallel-perf-output/time-range-0/cmd.txt:perf script --time=,9466.504461499 --ns
+ parallel-perf-output/time-range-1/cmd.txt:perf script --time=9466.504461500,9469.005396999 --ns
+ parallel-perf-output/time-range-2/cmd.txt:perf script --time=9469.005397000,9471.506332499 --ns
+ parallel-perf-output/time-range-3/cmd.txt:perf script --time=9471.506332500, --ns
+
+Any perf script command can be used. Standard output is redirected to files
+named out.txt in separate subdirectories under the output directory. Similarly,
+standard error is written to files named err.txt. In addition, files named
+cmd.txt contain the corresponding perf script command. After processing,
+err.txt files are removed if they are empty.
+
+If any job exits with a non-zero exit code, then all jobs are killed and no
+more are started. A message is printed if any job results in a non-empty
+err.txt file.
+
+There is a separate output subdirectory for each time range. If the --per-cpu
+option is used, these are further grouped under cpu-n subdirectories, e.g.
+
+ $ parallel-perf.py --per-cpu --nr=2 -- perf script --ns --cpu=0,1
+ All jobs finished successfully
+ $ tree parallel-perf-output
+ parallel-perf-output/
+ ├── cpu-0
+ │ ├── time-range-0
+ │ │ ├── cmd.txt
+ │ │ └── out.txt
+ │ └── time-range-1
+ │ ├── cmd.txt
+ │ └── out.txt
+ └── cpu-1
+ ├── time-range-0
+ │ ├── cmd.txt
+ │ └── out.txt
+ └── time-range-1
+ ├── cmd.txt
+ └── out.txt
+ $ find parallel-perf-output -name cmd.txt | sort | xargs grep -H .
+ parallel-perf-output/cpu-0/time-range-0/cmd.txt:perf script --cpu=0 --time=,9469.005396999 --ns
+ parallel-perf-output/cpu-0/time-range-1/cmd.txt:perf script --cpu=0 --time=9469.005397000, --ns
+ parallel-perf-output/cpu-1/time-range-0/cmd.txt:perf script --cpu=1 --time=,9469.005396999 --ns
+ parallel-perf-output/cpu-1/time-range-1/cmd.txt:perf script --cpu=1 --time=9469.005397000, --ns
+
+Subdivisions of time range, and cpus if the --per-cpu option is used, are
+expressed by the --time and --cpu perf script options respectively. If the
+supplied perf script command has a --time option, then that time range is
+subdivided, otherwise the time range given by 'time of first sample' to
+'time of last sample' is used (refer perf script --header-only). Similarly, the
+supplied perf script command may provide a --cpu option, and only those CPUs
+will be processed.
+
+Note there is special handling for processing Intel PT traces. If an interval is
+not specified and the perf record command contained the intel_pt event, then the
+time range will be subdivided in order to produce subdivisions that contain
+approximately the same amount of trace data. That is accomplished by counting
+double-quick (--itrace=qqi) samples, and choosing time ranges that encompass
+approximately the same number of samples. In that case, time ranges may not be
+the same for each CPU processed. For Intel PT, --per-cpu is the default, but
+that can be overridden by --no-per-cpu. Note, for Intel PT, double-quick
+decoding produces 1 sample for each PSB synchronization packet, which in turn
+come after a certain number of bytes output, determined by psb_period (refer
+perf Intel PT documentation). The minimum number of double-quick samples that
+will define a time range can be set by the --minimum option, which defaults to
+64.
+""")
+ ap.add_argument("-o", "--output", default="parallel-perf-output", help="output directory (default 'parallel-perf-output')")
+ 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)")
+ ap.add_argument("-n", "--nr", type=int, default=0, help="number of time subdivisions (default is the number of jobs)")
+ 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)")
+ ap.add_argument("-c", "--per-cpu", action="store_true", help="process data for each CPU in parallel")
+ ap.add_argument("-N", "--no-per-cpu", action="store_true", help="do not process data for each CPU in parallel")
+ ap.add_argument("-m", "--minimum", type=int, default=glb_min_samples, help="minimum data size (for Intel PT in PSBs)")
+ ap.add_argument("-D", "--dry-run", action="store_true", help="do not run any jobs, just show the perf script commands")
+ ap.add_argument("-q", "--quiet", action="store_true", help="do not print any messages except errors")
+ ap.add_argument("-v", "--verbose", action="store_true", help="print more messages")
+ ap.add_argument("-d", "--debug", action="store_true", help="print debugging messages")
+ cmd_line = list(args)
+ try:
+ split_pos = cmd_line.index("--")
+ cmd = cmd_line[split_pos + 1:]
+ args = cmd_line[:split_pos]
+ except:
+ cmd = None
+ args = cmd_line
+ a = ap.parse_args(args=args[1:])
+ v = Verbosity()
+ if a.quiet:
+ v.Quiet()
+ if a.verbose:
+ v.Verbose()
+ if a.debug:
+ v.Debug()
+ try:
+ if cmd == None:
+ if len(args) <= 1:
+ ap.print_help()
+ sys.exit(0)
+ raise Exception("Command line must contain '--' before perf command")
+ RunParallelPerf(a.output, cmd, a.jobs, a.nr, a.interval, a.per_cpu, a.no_per_cpu, a.minimum, a.dry_run, v)
+ except Exception as e:
+ print("Fatal error: ", str(e))
+ if a.debug:
+ raise
+ sys.exit(1)
+
+if __name__ == "__main__":
+ Main(sys.argv)
--
2.34.1
On Sun, Mar 10, 2024 at 09:35:02PM +0200, Adrian Hunter wrote:
> Add a Python script to run a perf script command multiple times in
> parallel, using perf script options --cpu and --time so that each job
> processes a different chunk of the data.
>
> Refer to the script's own help text at the end of the patch for more
> details.
>
> The script is useful for Intel PT traces, that can be efficiently
> decoded by perf script when split by CPU and/or time ranges. Running
> jobs in parallel can decrease the overall decoding time.
This only optimizes for the run time of the decoder. Often when you do
analysis you have a non trivial part of it in some analysis script too,
but you currently have no directi / easy way to paralelize that. It would
be better to support parallel pipelines.
TBH I'm not sure the script is worth it. If you need to do parallel
pipelines (which imho is the common case) it's probably better to just
write a custom shell script, which is not that difficult. It might be
better to have a helper that makes writing such scripts easier,
e.g. figuring out reasonable options for manual parallelization
based on the input file. I think parts of your script do that, maybe
it is usable for that.
Also as a default output it would be better to just merge the
original output in order and output it on stdout.
You should probably limit the number of jobs to some minimum
length, otherwise on systems with many CPUs there might be
inefficiently short jobs.
-Andi
On 11/03/24 18:13, Andi Kleen wrote:
> On Sun, Mar 10, 2024 at 09:35:02PM +0200, Adrian Hunter wrote:
>> Add a Python script to run a perf script command multiple times in
>> parallel, using perf script options --cpu and --time so that each job
>> processes a different chunk of the data.
>>
>> Refer to the script's own help text at the end of the patch for more
>> details.
>>
>> The script is useful for Intel PT traces, that can be efficiently
>> decoded by perf script when split by CPU and/or time ranges. Running
>> jobs in parallel can decrease the overall decoding time.
>
> This only optimizes for the run time of the decoder. Often when you do
> analysis you have a non trivial part of it in some analysis script too,
> but you currently have no directi / easy way to paralelize that. It would
> be better to support parallel pipelines.
It will parallelize any scripts and / or dlfilters that perf script
itself executes.
>
> TBH I'm not sure the script is worth it. If you need to do parallel
> pipelines (which imho is the common case) it's probably better to just
> write a custom shell script, which is not that difficult.
It can be a pain to figure out how best to split the data if it is not
evenly distributed.
The script also has value as a reference or starting point for
users.
> It might be
> better to have a helper that makes writing such scripts easier,
> e.g. figuring out reasonable options for manual parallelization
> based on the input file. I think parts of your script do that, maybe
> it is usable for that.
The --dry-run option shows the perf script commands, but an option
to pipe through another command could be added.
>
> Also as a default output it would be better to just merge the
> original output in order and output it on stdout.
That assumes that the output comes from perf script printf
output and not a perf script _script_.
If the data is split by CPU, it will not be in time order
if it is simply concatenated back together.
>
> You should probably limit the number of jobs to some minimum
> length, otherwise on systems with many CPUs there might be
> inefficiently short jobs.
That happens for Intel PT (64 PSB minimum), but could be added
for the normal case also.