# -*- test-case-name: epsilon.test.test_benchmark -*-
"""
Functions for running a Python file in a child process and recording resource
usage information and other statistics about it.
"""
import os, time, sys, socket, StringIO, pprint, errno
import twisted
from twisted.python import log, filepath, failure, util
from twisted.internet import reactor, protocol, error, defer
from twisted.protocols import policies
import epsilon
from epsilon import structlike
from epsilon import juice
from epsilon.test import utils
class diskstat(structlike.record(
'readCount mergedReadCount readSectorCount readMilliseconds '
'writeCount mergedWriteCount writeSectorCount writeMilliseconds '
'outstandingIOCount ioMilliseconds weightedIOMilliseconds')):
"""
Represent the I/O stats of a single device, as reported by Linux's disk
stats.
"""
class partitionstat(structlike.record(
'readCount readSectorCount writeCount writeSectorCount')):
"""
Like diskstat, but for a partition. Less information is made available by
Linux for partitions, so this has fewer attributes.
"""
def parseDiskStatLine(L):
"""
Parse a single line from C{/proc/diskstats} into a two-tuple of the name of
the device to which it corresponds (ie 'hda') and an instance of the
appropriate record type (either L{partitionstat} or L{diskstat}).
"""
parts = L.split()
device = parts[2]
if len(parts) == 7:
factory = partitionstat
else:
factory = diskstat
return device, factory(*map(int, parts[3:]))
def parseDiskStats(fObj):
"""
Parse a file-like object containing lines formatted like those in
C{/proc/diskstats}. Yield two-tuples of information for each line.
"""
for L in fObj:
yield parseDiskStatLine(L)
def captureStats():
"""
Parse the current contents of C{/proc/diskstats} into a dict mapping device
names to instances of the appropriate stat record.
"""
return dict(parseDiskStats(file('/proc/diskstats')))
class ResourceSnapshot(structlike.record('time disk partition size')):
"""
Represents the state of some resources on the system at a particular moment
in time.
@ivar time: The time at which the stats associated with this instance were
recorded.
@ivar disk: A C{diskstat} instance created from stats available that the
given time.
@ivar partition: A C{diskstat} instance created from stats available that
the given time.
@ivar size: Total size of all files beneath a particular directory.
"""
class ProcessDied(Exception):
"""
Encapsulates process state and failure mode.
"""
def __init__(self, exitCode, signal, status, output):
self.exitCode = exitCode
self.signal = signal
self.status = status
self.output = output
Exception.__init__(self)
class BasicProcess(protocol.ProcessProtocol, policies.TimeoutMixin):
"""
The simplest possible process protocol. It doesn't do anything except what
is absolutely mandatory of any conceivable ProcessProtocol.
"""
timedOut = False
BACKCHANNEL_OUT = 3
BACKCHANNEL_IN = 4
def __init__(self, whenFinished, path):
self.whenFinished = whenFinished
self.path = path
self.output = []
def connectionMade(self):
self.setTimeout(900.0)
def timeoutConnection(self):
self.timedOut = True
self.transport.signalProcess('KILL')
def childDataReceived(self, childFD, data):
self.resetTimeout()
self.output.append((childFD, data))
def childConnectionLost(self, childFD):
self.resetTimeout()
self.output.append((childFD, None))
def processEnded(self, reason):
# XXX Okay, I'm a liar. This doesn't do everything. Strictly speaking
# we shouldn't fire completion notification until the process has
# terminated *and* the file descriptors have all been closed. We're
# not supporting passing file descriptors from the child to a
# grandchild here, though. Don't Do It.
d, self.whenFinished = self.whenFinished, None
o, self.output = self.output, None
if reason.check(error.ProcessDone):
d.callback((self, reason.value.status, o))
elif self.timedOut:
d.errback(error.TimeoutError())
elif reason.check(error.ProcessTerminated):
d.errback(failure.Failure(ProcessDied(
reason.value.exitCode,
reason.value.signal,
reason.value.status, o)))
else:
d.errback(reason.value)
self.setTimeout(None)
def spawn(cls, executable, args, path, env, spawnProcess=None):
"""
Run an executable with some arguments in the given working directory with
the given environment variables.
Returns a Deferred which fires with a two-tuple of (exit status, output
list) if the process terminates without timing out or being killed by a
signal. Otherwise, the Deferred errbacks with either L{error.TimeoutError}
if any 10 minute period passes with no events or L{ProcessDied} if it is
killed by a signal.
On success, the output list is of two-tuples of (file descriptor, bytes).
"""
d = defer.Deferred()
proto = cls(d, filepath.FilePath(path))
if spawnProcess is None:
spawnProcess = reactor.spawnProcess
spawnProcess(
proto,
executable,
[executable] + args,
path=path,
env=env,
childFDs={0: 'w', 1: 'r', 2: 'r',
cls.BACKCHANNEL_OUT: 'r',
cls.BACKCHANNEL_IN: 'w'})
return d
spawn = classmethod(spawn)
class Change(object):
"""
Stores two ResourceSnapshots taken at two different times.
"""
def start(self, path, disk, partition):
# Do these three things as explicit, separate statments to make sure
# gathering disk stats isn't accidentally included in the duration.
startSize = getSize(path)
beforeDiskStats = captureStats()
startTime = time.time()
self.before = ResourceSnapshot(
time=startTime,
disk=beforeDiskStats.get(disk, None),
partition=beforeDiskStats.get(partition, None),
size=startSize)
def stop(self, path, disk, partition):
# Do these three things as explicit, separate statments to make sure
# gathering disk stats isn't accidentally included in the duration.
endTime = time.time()
afterDiskStats = captureStats()
endSize = getSize(path)
self.after = ResourceSnapshot(
time=endTime,
disk=afterDiskStats.get(disk, None),
partition=afterDiskStats.get(partition, None),
size=endSize)
class BenchmarkProcess(BasicProcess):
START = '\0'
STOP = '\1'
def __init__(self, *a, **kw):
BasicProcess.__init__(self, *a, **kw)
# Figure out where the process is running.
self.partition = discoverCurrentWorkingDevice().split('/')[-1]
self.disk = self.partition.rstrip('0123456789')
# Keep track of stats for the entire process run.
self.overallChange = Change()
self.overallChange.start(self.path, self.disk, self.partition)
# Just keep track of stats between START and STOP events.
self.benchmarkChange = Change()
def connectionMade(self):
return BasicProcess.connectionMade(self)
def startTiming(self):
self.benchmarkChange.start(self.path, self.disk, self.partition)
self.transport.writeToChild(self.BACKCHANNEL_IN, self.START)
def stopTiming(self):
self.benchmarkChange.stop(self.path, self.disk, self.partition)
self.transport.writeToChild(self.BACKCHANNEL_IN, self.STOP)
def childDataReceived(self, childFD, data):
if childFD == self.BACKCHANNEL_OUT:
self.resetTimeout()
for byte in data:
if byte == self.START:
self.startTiming()
elif byte == self.STOP:
self.stopTiming()
else:
self.transport.signalProcess('QUIT')
else:
return BasicProcess.childDataReceived(self, childFD, data)
def processEnded(self, reason):
self.overallChange.stop(self.path, self.disk, self.partition)
return BasicProcess.processEnded(self, reason)
STATS_VERSION = 0
class Results(juice.Command):
commandName = 'Result'
arguments = [
# Stats version - change this whenever the meaning of something changes
# or a field is added or removed.
('version', juice.Integer()),
# If an error occurred while collecting these stats - this probably
# means they're bogus.
('error', juice.Boolean()),
# If a particular timeout (See BasicProcess.connectionMade) elapsed
# with no events whatsoever from the benchmark process.
('timeout', juice.Boolean()),
# A unique name identifying the benchmark for which these are stats.
('name', juice.Unicode()),
# The name of the benchmark associated with these stats.
('host', juice.Unicode()),
# The sector size of the disk on which these stats were collected
# (sectors are a gross lie, this is really the block size, and
# everything else that talks about sectors is really talking about
# blocks).
('sector_size', juice.Integer()),
# Hex version info for the Python which generated these stats.
('python_version', juice.Unicode()),
# Twisted SVN revision number used to generate these stats.
('twisted_version', juice.Unicode()),
# Divmod SVN revision number used to generate these stats.
('divmod_version', juice.Unicode()),
# Number of seconds between process startup and termination.
('elapsed', juice.Float()),
# Size, in bytes, of the directory in which the child process was run.
('filesystem_growth', juice.Integer()),
# Number of reads issued on the partition over the lifetime of the
# child process. This may include reads from other processes, if any
# were active on the same disk when the stats were collected.
('read_count', juice.Integer(optional=True)),
# Number of sectors which were read from the partition over the
# lifetime of the child process. Same caveat as above.
('read_sectors', juice.Integer(optional=True)),
# Number of writes issued to the partition over the lifetime of the
# child process. Same caveat as above.
('write_count', juice.Integer(optional=True)),
# Number of sectors which were written to the partition over the
# lifetime of the child process. Same caveat as above.
('write_sectors', juice.Integer(optional=True)),
# Number of milliseconds spent blocked on reading from the disk over
# the lifetime of the child process. Same caveat as above.
('read_ms', juice.Integer(optional=True)),
# Number of milliseconds spent blocked on writing to the disk over the
# lifetime of the child process. Same caveat as above.
('write_ms', juice.Integer(optional=True)),
]
hostname = socket.gethostname()
assert hostname != 'localhost', "Fix your computro."
def formatResults(name, sectorSize, before, after, error, timeout):
output = StringIO.StringIO()
jj = juice.Juice(issueGreeting=False)
tt = utils.FileWrapper(output)
jj.makeConnection(tt)
if after.partition is not None:
read_count = after.partition.readCount - before.partition.readCount
read_sectors = after.partition.readSectorCount - before.partition.readSectorCount
write_count = after.partition.writeCount - before.partition.writeCount
write_sectors = after.partition.writeSectorCount - before.partition.writeSectorCount
else:
read_count = None
read_sectors = None
write_count = None
write_sectors = None
if after.disk is not None:
read_ms = after.disk.readMilliseconds - before.disk.readMilliseconds
write_ms = after.disk.writeMilliseconds - before.disk.writeMilliseconds
else:
read_ms = None
write_ms = None
twisted_version = twisted.version._getSVNVersion()
if twisted_version is None:
twisted_version = twisted.version.short()
epsilon_version = epsilon.version._getSVNVersion()
if epsilon_version is None:
epsilon_version = epsilon.version.short()
Results(
version=STATS_VERSION,
error=error,
timeout=timeout,
name=name,
host=hostname,
elapsed=after.time - before.time,
sector_size=sectorSize,
read_count=read_count,
read_sectors=read_sectors,
read_ms=read_ms,
write_count=write_count,
write_sectors=write_sectors,
write_ms=write_ms,
filesystem_growth=after.size - before.size,
python_version=unicode(sys.hexversion),
twisted_version=twisted_version,
divmod_version=epsilon_version,
).do(jj, requiresAnswer=False)
return output.getvalue()
def reportResults(results):
print results
print
fObj = file('output', 'ab')
fObj.write(results)
fObj.close()
def discoverCurrentWorkingDevice():
"""
Return a short string naming the device which backs the current working
directory, ie C{/dev/hda1}.
"""
possibilities = []
cwd = os.getcwd()
for L in file('/etc/mtab'):
parts = L.split()
if cwd.startswith(parts[1]):
possibilities.append((len(parts[1]), parts[0]))
possibilities.sort()
return possibilities[-1][-1]
def getSize(p):
"""
@type p: L{twisted.python.filepath.FilePath}
@return: The size, in bytes, of the given path and all its children.
"""
result = 0
for ch in p.walk():
try:
result += ch.getsize()
except OSError, e:
if e.errno == errno.ENOENT:
# XXX FilePath is broken
if os.path.islink(ch.path):
result += len(os.readlink(ch.path))
else:
raise
else:
raise
return result
def getSectorSize(p):
return os.statvfs(p.path).f_bsize
def _bench(name, workingPath, function):
d = function()
def later(result):
err = timeout = False
if isinstance(result, failure.Failure):
err = True
if result.check(error.TimeoutError):
log.msg("Failing because timeout!")
timeout = True
elif result.check(ProcessDied):
log.msg("Failing because Failure!")
pprint.pprint(result.value.output)
print result.value.exitCode, result.value.signal
else:
log.err(result)
else:
proto, status, output = result
stderr = [bytes for (fd, bytes) in output if fd == 2]
if status or stderr != [None]:
err = True
log.msg("Failing because stderr or bad status")
pprint.pprint(result)
for n, change in [(name + '-overall', proto.overallChange),
(name + '-benchmark', proto.benchmarkChange)]:
reportResults(formatResults(
n,
getSectorSize(workingPath),
change.before,
change.after,
err,
timeout))
return d.addBoth(later)
def bench(name, path, func):
log.startLogging(sys.stdout)
log.msg("Running " + name)
d = _bench(name, path, func)
d.addErrback(log.err)
d.addCallback(lambda ign: reactor.stop())
reactor.run()
def makeBenchmarkRunner(path, args):
"""
Make a function that will run two Python processes serially: first one
which calls the setup function from the given file, then one which calls
the execute function from the given file.
"""
def runner():
return BenchmarkProcess.spawn(
executable=sys.executable,
args=['-Wignore'] + args,
path=path.path,
env=os.environ)
return runner
def start():
"""
Start recording stats. Call this from a benchmark script when your setup
is done. Call this at most once.
@raise RuntimeError: Raised if the parent process responds with anything
other than an acknowledgement of this message.
"""
os.write(BenchmarkProcess.BACKCHANNEL_OUT, BenchmarkProcess.START)
response = util.untilConcludes(os.read, BenchmarkProcess.BACKCHANNEL_IN, 1)
if response != BenchmarkProcess.START:
raise RuntimeError(
"Parent process responded with %r instead of START " % (response,))
def stop():
"""
Stop recording stats. Call this from a benchmark script when the code you
want benchmarked has finished. Call this exactly the same number of times
you call L{start} and only after calling it.
@raise RuntimeError: Raised if the parent process responds with anything
other than an acknowledgement of this message.
"""
os.write(BenchmarkProcess.BACKCHANNEL_OUT, BenchmarkProcess.STOP)
response = util.untilConcludes(os.read, BenchmarkProcess.BACKCHANNEL_IN, 1)
if response != BenchmarkProcess.STOP:
raise RuntimeError(
"Parent process responded with %r instead of STOP" % (response,))
def main():
"""
Run me with the filename of a benchmark script as an argument. I will time
it and append the results to a file named output in the current working
directory.
"""
name = sys.argv[1]
path = filepath.FilePath('.stat').temporarySibling()
path.makedirs()
func = makeBenchmarkRunner(path, sys.argv[1:])
try:
bench(name, path, func)
finally:
path.remove()
if __name__ == '__main__':
main()
syntax highlighted by Code2HTML, v. 0.9.1