#!/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
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 [ f"Non-empty error file {self.stderr_name}" ] return []
def TidyUp(self):
self.RemoveEmptyErrFile()
def RawPollWait(self, p, wait): if wait: return p.wait() return p.poll()
def Poll(self, wait=False): ifnot self.popen: returnNone
result = self.RawPollWait(self.popen, wait) if self.consumer:
res = result
result = self.RawPollWait(self.consumer, wait) if result != Noneand res == None:
self.popen.kill()
result = None elif result == 0 and res != Noneand res != 0:
result = res if result != None:
self.TidyUp() return result
def Wait(self): return self.Poll(wait=True)
def Kill(self): ifnot self.popen: return
self.popen.kill() if self.consumer:
self.consumer.kill()
def KillWork(worklist, verbosity): for w in worklist:
w.Kill() for w in worklist:
w.Wait()
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 whileTrue:
name = f"{orig_name} {nr}" if name notin result: break
nr += 1
result[name] = value return result
def HeaderField(hdr_dict, hdr_fld): if hdr_fld notin hdr_dict: raise Exception(f"'{hdr_fld}' missing from header information") return hdr_dict[hdr_fld]
# Represent the position of an option within a command string # and provide the option value and/or remove the option 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] == f"-{short_name}": if len(opt) == 2: if pos + 1 < len(args):
self.Init(pos, pos + 1, 0, 0) else:
self.Init(error = f"-{short_name} option missing value") else:
self.Init(pos, pos, 0, 2) return if opt[:n] == f"--{long_name}": if len(opt) == n: if pos + 1 < len(args):
self.Init(pos, pos + 1, 0, 0) else:
self.Init(error = f"--{long_name} option missing value") elif opt[n] == "=":
self.Init(pos, pos, 0, n + 1) else:
self.Init(error = f"--{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 = f"-{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(f"perf command {p.error}")
file_name = p.Value() ifnot os.path.exists(file_name): raise Exception(f"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(f"{err_prefix}{p.error}")
value = p.Value() if remove:
p.Remove(args) return value
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) whileTrue:
line = p.stdout.readline() if line:
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
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 min_size < 1:
min_size = 1
if mx < min_size: # 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, min_interval) ]
if nr:
divisor = nr
min_size = 1 else:
divisor = NumberOfCPUs()
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, min_interval): if n <= len(time_ranges): return time_ranges
duration = TotalDuration(time_ranges)
interval = duration / n if interval < min_interval:
interval = min_interval 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(f"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(f"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(f"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 == Noneor 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 == Noneor 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, a): for arg_name in vars(a):
setattr(self, arg_name, getattr(a, arg_name))
self.orig_nr = self.nr
self.orig_cmd = list(self.cmd)
self.perf = self.cmd[0] if os.path.exists(self.output_dir): raise Exception(f"Output '{self.output_dir}' already exists") if self.jobs < 0 or self.nr < 0 or self.interval < 0: raise Exception("Bad options (negative values): try -h option for help") if self.nr != 0 and self.interval != 0: raise Exception("Cannot specify number of time subdivisions and time interval") if self.jobs == 0:
self.jobs = NumberOfCPUs() if self.nr == 0 and self.interval == 0: if self.per_cpu:
self.nr = 1 else:
self.nr = self.jobs
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 DefaultToPerCPU(self): # --no-per-cpu option takes precedence if self.no_per_cpu: returnFalse ifnot self.PerfDataRecordedPerCPU(): returnFalse # 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(): returnTrue returnFalse
def Config(self):
self.Init()
self.ExtractTimeInfo() ifnot 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()) returnTrue
result = RunWork(self.worklist, self.jobs, verbosity=self.verbosity) if self.verbosity.verbose:
print(glb_prog_name, "done") return result
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.
Any perf script command can be used, including the use of perf script options
--dlfilter and --script, so that the benefit of running parallel jobs
naturally extends to them also.
If option --pipe-to is used, standard output is first piped through that
command. Beware, if the command fails (e.g. grep with no matches), it will be
considered a fatal error.
Final 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.
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.
To prevent time intervals becoming too small, the --min-interval option can
be used.
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 --min_size option, which defaults to
64. """)
ap.add_argument("-o", "--output-dir", 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("-m", "--min-interval", type=float, default=glb_min_interval, help=f"minimum interval (default {glb_min_interval} seconds)")
ap.add_argument("-p", "--pipe-to", help="command to pipe output to (optional)")
ap.add_argument("-N", "--no-per-cpu", action="store_true", help="do not process data for each CPU in parallel")
ap.add_argument("-b", "--min_size", 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:])
a.cmd = cmd
a.verbosity = Verbosity(a.quiet, a.verbose, a.debug) try: if a.cmd == None: if len(args) <= 1:
ap.print_help() returnTrue raise Exception("Command line must contain '--' before perf command") return RunParallelPerf(a) except Exception as e:
print("Fatal error: ", str(e)) if a.debug: raise returnFalse
if __name__ == "__main__": ifnot Main(sys.argv):
sys.exit(1)
Messung V0.5
¤ Dauer der Verarbeitung: 0.14 Sekunden
(vorverarbeitet)
¤
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.