def create_worker_list(prm):
# for each thread set up FsDriftWorkload instance,
# create a thread instance, and delete the thread-ready file
thread_list = []
for k in range(0, prm.threads):
nextinv = worker_thread.FsDriftWorkload(prm)
nextinv.tid = '%02d' % k
t = invoke_process.subprocess(nextinv)
thread_list.append(t)
ensure_deleted(nextinv.gen_thread_ready_fname(nextinv.tid))
return thread_list
def run(self):
ensure_deleted(self.launch_fn)
ensure_deleted(self.pickle_fn)
with open(self.launch_fn, 'w') as launch_file:
launch_file.write(self.remote_cmd)
launch_file.close()
self.log.debug('waiting for pickle file %s' % self.pickle_fn)
self.status = NOTOK # premature exit means failure
while not os.path.exists(self.pickle_fn):
if os.path.exists(self.params.abort_path):
self.log.info('test abort seen by host ' + self.remote_host)
return
time.sleep(2)
self.status = OK # success!
def run_multi_host_workload(prm, log):
# construct list of ssh threads to invoke in parallel
if os.getenv('PYPY'):
python_prog = os.getenv('PYPY')
elif sys.version.startswith('2'):
python_prog = 'python'
elif sys.version.startswith('3'):
python_prog = 'python3'
else:
raise Exception('unrecognized python version %s' % sys.version)
log.debug('python_prog = %s' % python_prog)
remote_thread_list = []
host_ct = len(prm.host_set)
for j in range(0, len(prm.host_set)):
remote_host = prm.host_set[j]
fsd_remote_pgm = os.path.join(prm.fsd_remote_dir, 'fs-drift-remote.py')
this_remote_cmd = '%s %s --network-sync-dir %s ' \
% (prm.python_prog, fsd_remote_pgm, prm.network_shared_path)
this_remote_cmd += ' --as-host %s' % remote_host
log.debug(this_remote_cmd)
if prm.launch_as_daemon:
remote_thread_list.append(
launcher_thread.launcher_thread(prm, log, remote_host,
this_remote_cmd))
else:
remote_thread_list.append(
ssh_thread.ssh_thread(log, remote_host, this_remote_cmd))
# start them, pacing starts so that we don't get ssh errors
for t in remote_thread_list:
if prm.launch_as_daemon:
time.sleep(0.1)
t.start()
# wait for hosts to arrive at starting gate
# if only one host, then no wait will occur
# as starting gate file is already present
# every second we resume scan from last host file not found
exception_seen = None
abortfn = prm.abort_path
sec_delta = 0.5
# timeout if no host replies in next host_timeout seconds
per_host_timeout = 10.0
all_host_timeout = 5.0 + len(prm.host_set) / 3
if all_host_timeout < per_host_timeout:
per_host_timeout = all_host_timeout / 2
hosts_ready = False # set scope outside while loop
last_host_seen = -1
sec = 0.0
start_loop_start = time.time()
try:
while sec < per_host_timeout:
# HACK to force directory entry coherency for Gluster
#ndirlist = os.listdir(prm.network_shared_path)
#log.debug('shared dir list: ' + str(ndirlist))
hosts_ready = True
if os.path.exists(abortfn):
raise FsDriftException('worker host signaled abort')
for j in range(last_host_seen + 1, len(prm.host_set)):
h = prm.host_set[j]
fn = multi_thread_workload.gen_host_ready_fname(prm, h.strip())
log.debug('checking for host filename ' + fn)
if not os.path.exists(fn):
log.info('did not see host filename %s after %f sec' %
(fn, sec))
hosts_ready = False
break
log.debug('saw host filename ' + fn)
last_host_seen = j # saw this host's ready file
# we exit while loop only if no hosts in per_host_timeout seconds
sec = 0.0
if hosts_ready:
break
# if one of ssh threads has died, no reason to continue
kill_remaining_threads = False
for t in remote_thread_list:
if not t.is_alive():
log.error('thread %s has died' % t)
kill_remaining_threads = True
break
if kill_remaining_threads:
break
# be patient for large tests
# give user some feedback about
# how many hosts have arrived at the starting gate
time.sleep(sec_delta)
sec += sec_delta
time_since_loop_start = time.time() - start_loop_start
log.debug('last_host_seen=%d sec=%d' % (last_host_seen, sec))
if time_since_loop_start > all_host_timeout:
kill_remaining_threads = True
break
except KeyboardInterrupt as e:
log.error('saw SIGINT signal, aborting test')
exception_seen = e
except Exception as e:
exception_seen = e
log.exception(e)
hosts_ready = False
if not hosts_ready:
multi_thread_workload.abort_test(prm.abort_path, remote_thread_list)
if not exception_seen:
log.info(
'no additional hosts reached starting gate within %5.1f seconds'
% per_host_timeout)
return NOTOK
else:
raise exception_seen
else:
# ask all hosts to start the test
# this is like firing the gun at the track meet
try:
sync_files.write_sync_file(prm.starting_gun_path, 'hi')
log.debug('starting all threads by creating starting gun file %s' %
prm.starting_gun_path)
except IOError as e:
log.error('error writing starting gun file: %s' %
os.strerror(e.errno))
multi_thread_workload.abort_test(prm.abort_path,
remote_thread_list)
raise e
# wait for them to finish
for t in remote_thread_list:
t.join()
if t.status != OK:
log.error('ssh thread for host %s completed with status %d' %
(t.remote_host, t.status))
# attempt to aggregate results by reading pickle files
# containing SmallfileWorkload instances
# with counters and times that we need
try:
invoke_list = []
one_shot_delay = True
for h in prm.host_set: # for each host in test
# read results for each thread run in that host
# from python pickle of the list of SmallfileWorkload objects
pickle_fn = multi_thread_workload.host_result_filename(prm, h)
log.debug('reading pickle file: %s' % pickle_fn)
host_invoke_list = []
try:
if one_shot_delay and not os.path.exists(pickle_fn):
# all threads have joined already, they are done
# we allow > 1 sec
# for this (NFS) client to see other clients' files
time.sleep(1.2)
one_shot_delay = False
host_invoke_list = read_pickle(pickle_fn)
log.debug(' read %d invoke objects' % len(host_invoke_list))
invoke_list.extend(host_invoke_list)
ensure_deleted(pickle_fn)
except IOError as e:
if e.errno != errno.ENOENT:
raise e
log.error(' pickle file %s not found' % pickle_fn)
output_results.output_results(prm, invoke_list)
except IOError as e:
log.exception(e)
log.error('host %s filename %s: %s' % (h, pickle_fn, str(e)))
return NOTOK
except KeyboardInterrupt as e:
log.error('control-C signal seen (SIGINT)')
return NOTOK
except FsDriftException as e:
log.exception(e)
return NOTOK
return (OK)
def run_multi_thread_workload(prm):
host = prm.as_host
if host == None:
host = 'localhost'
prm_slave = (prm.host_set != [])
# FIXME: get coherent logging level interface
verbose = os.getenv('LOGLEVEL_DEBUG' != None)
host_startup_timeout = 5 + len(prm.host_set) / 3
# for each thread set up SmallfileWorkload instance,
# create a thread instance, and delete the thread-ready file
thread_list = create_worker_list(prm)
my_host_invoke = thread_list[0].invoke
my_log = fsd_log.start_log('%s.master' % host)
my_log.debug(prm)
# start threads, wait for them to reach starting gate
# to do this, look for thread-ready files
for t in thread_list:
ensure_deleted(t.invoke.gen_thread_ready_fname(t.invoke.tid))
for t in thread_list:
t.start()
my_log.debug('started %d worker threads on host %s' %
(len(thread_list), host))
# wait for all threads to reach the starting gate
# this makes it more likely that they will start simultaneously
abort_fname = prm.abort_path
thread_count = len(thread_list)
thread_to_wait_for = 0
startup_timeout = 3
sec = 0.0
while sec < startup_timeout:
for k in range(thread_to_wait_for, thread_count):
t = thread_list[k]
fn = t.invoke.gen_thread_ready_fname(t.invoke.tid)
if not os.path.exists(fn):
my_log.debug('thread %d thread-ready file %s not found yet with %f sec left' %
(k, fn, (startup_timeout - sec)))
break
thread_to_wait_for = k + 1
# we only timeout if no more threads have reached starting gate
# in startup_timeout sec
sec = 0.0
if thread_to_wait_for == thread_count:
break
if os.path.exists(abort_fname):
break
sec += 0.5
time.sleep(0.5)
# if all threads didn't make it to the starting gate
if thread_to_wait_for < thread_count:
abort_test(abort_fname, thread_list)
raise FsDriftException('only %d threads reached starting gate'
% thread_to_wait_for)
# declare that this host is at the starting gate
if prm_slave:
host_ready_fn = gen_host_ready_fname(prm, prm.as_host)
my_log.debug('host %s creating ready file %s' %
(my_host_invoke.onhost, host_ready_fn))
common.touch(host_ready_fn)
sg = prm.starting_gun_path
if not prm_slave:
my_log.debug('wrote starting gate file ')
sync_files.write_sync_file(sg, 'hi there')
# wait for starting_gate file to be created by test driver
# every second we resume scan from last host file not found
if prm_slave:
my_log.debug('awaiting ' + sg)
for sec in range(0, host_startup_timeout+3):
# hack to ensure that directory is up to date
# ndlist = os.listdir(my_host_invoke.network_dir)
# if verbose: print(str(ndlist))
if os.path.exists(sg):
break
if os.path.exists(prm.abort_path):
log.info('saw abort file %s, aborting test' % prm.abort_path)
break
time.sleep(1)
if not os.path.exists(sg):
abort_test(prm.abort_path, thread_list)
raise Exception('starting signal not seen within %d seconds'
% host_startup_timeout)
if verbose:
print('starting test on host ' + host + ' in 2 seconds')
time.sleep(2 + random.random()) # let other hosts see starting gate file
# FIXME: don't timeout the test,
# instead check thread progress and abort if you see any of them stalled
# but if servers are heavily loaded you can't rely on filesystem
# wait for all threads on this host to finish
for t in thread_list:
my_log.debug('waiting for thread %s' % t.invoke.tid)
t.retrieve()
t.join()
# if not a slave of some other host, print results (for this host)
if not prm_slave:
try:
worker_list = [ t.invoke for t in thread_list ]
output_results.output_results(prm, worker_list)
except FsDriftException as e:
print('ERROR: ' + str(e))
return NOTOK
else:
# if we are participating in a multi-host test
# then write out this host's result in pickle format
# so test driver can pick up result
result_filename = host_result_filename(prm, prm.as_host)
my_log.debug('saving result to filename %s' % result_filename)
worker_list = [ t.invoke for t in thread_list ]
sync_files.write_pickle(result_filename, worker_list)
time.sleep(1.2) # for benefit of NFS with actimeo=1
return OK