def __init__(self, db_name='fishtest_new'):
# MongoDB server is assumed to be on the same machine, if not user should use
# ssh with port forwarding to access the remote host.
self.conn = MongoClient(os.getenv('FISHTEST_HOST') or 'localhost')
self.db = self.conn[db_name]
self.userdb = UserDb(self.db)
self.actiondb = ActionDb(self.db)
self.runs = self.db['runs']
self.chunk_size = 1000
class RunDb:
def __init__(self, db_name='fishtest_new'):
# MongoDB server is assumed to be on the same machine, if not user should use
# ssh with port forwarding to access the remote host.
self.conn = MongoClient(os.getenv('FISHTEST_HOST') or 'localhost')
self.db = self.conn[db_name]
self.userdb = UserDb(self.db)
self.actiondb = ActionDb(self.db)
self.regressiondb = RegressionDb(self.db)
self.runs = self.db['runs']
self.chunk_size = 1000
def build_indices(self):
self.runs.ensure_index([('finished', ASCENDING), ('last_updated', DESCENDING)])
def generate_tasks(self, num_games):
tasks = []
remaining = num_games
while remaining > 0:
task_size = min(self.chunk_size, remaining)
tasks.append({
'num_games': task_size,
'pending': True,
'active': False,
})
remaining -= task_size
return tasks
def new_run(self, base_tag, new_tag, num_games, tc, book, book_depth, threads, base_options, new_options,
info='',
resolved_base='',
resolved_new='',
msg_base='',
msg_new='',
base_signature='',
new_signature='',
regression_test=False,
start_time=None,
sprt=None,
spsa=None,
username=None,
tests_repo=None,
throughput=1000,
priority=0):
if start_time == None:
start_time = datetime.utcnow()
run_args = {
'base_tag': base_tag,
'new_tag': new_tag,
'num_games': num_games,
'tc': tc,
'book': book,
'book_depth': book_depth,
'threads': threads,
'regression_test': regression_test,
'resolved_base': resolved_base,
'resolved_new': resolved_new,
'msg_base': msg_base,
'msg_new': msg_new,
'base_options': base_options,
'new_options': new_options,
'info': info,
'base_signature': base_signature,
'new_signature': new_signature,
'username': username,
'tests_repo': tests_repo,
'throughput': throughput,
'priority': priority,
'internal_priority': - time.mktime(start_time.timetuple()),
}
if sprt != None:
run_args['sprt'] = sprt
if spsa != None:
run_args['spsa'] = spsa
new_run = {
'args': run_args,
'start_time': start_time,
'last_updated': start_time,
# Will be filled in by tasks, indexed by task-id
'tasks': self.generate_tasks(num_games),
# Aggregated results
'results': { 'wins': 0, 'losses': 0, 'draws': 0 },
'results_stale': False,
'finished': False,
'approved': False,
'approver': '',
}
# Check for an existing approval matching the git commit SHAs
def get_approval(sha):
q = { '$or': [{ 'args.resolved_base': sha }, { 'args.resolved_new': sha }], 'approved': True }
return self.runs.find_one(q)
base_approval = get_approval(resolved_base)
new_approval = get_approval(resolved_new)
allow_auto = username in ['mcostalba', 'jkiiski', 'glinscott', 'lbraesch']
if base_approval != None and new_approval != None and allow_auto:
new_run['approved'] = True
new_run['approver'] = new_approval['approver']
return self.runs.insert(new_run)
def get_machines(self):
machines = []
for run in self.runs.find({'tasks': {'$elemMatch': {'active': True}}}):
for task in run['tasks']:
if task['active']:
machine = copy.copy(task['worker_info'])
machine['last_updated'] = task.get('last_updated', None)
machine['run'] = run
machine['nps'] = task.get('nps', 0)
# TODO(glinscott): Temporary - remove once worker version >= 41
if not isinstance(machine['uname'], basestring):
machine['uname'] = machine['uname'][0] + machine['uname'][2]
machines.append(machine)
return machines
def get_run(self, id):
return self.runs.find_one({'_id': ObjectId(id)})
def get_run_to_build(self):
return self.runs.find_one({'binaries_url': {'$exists': False}, 'finished': False, 'deleted': {'$exists': False}})
def get_runs(self):
return list(self.get_unfinished_runs()) + self.get_finished_runs()[0]
def get_unfinished_runs(self):
return self.runs.find({'finished': False},
sort=[('last_updated', DESCENDING), ('start_time', DESCENDING)])
def get_finished_runs(self, skip=0, limit=0, username=''):
q = {'finished': True, 'deleted': {'$exists': False}}
if len(username) > 0:
q['args.username'] = username
c = self.runs.find(q, skip=skip, limit=limit, sort=[('last_updated', DESCENDING)])
return (list(c), c.count())
def get_results(self, run):
if not run['results_stale']:
return run['results']
results = { 'wins': 0, 'losses': 0, 'draws': 0, 'crashes': 0, 'time_losses':0 }
for task in run['tasks']:
if 'stats' in task:
stats = task['stats']
results['wins'] += stats['wins']
results['losses'] += stats['losses']
results['draws'] += stats['draws']
results['crashes'] += stats['crashes']
results['time_losses'] += stats.get('time_losses', 0)
if 'sprt' in run['args'] and 'state' in run['args']['sprt']:
results['sprt'] = run['args']['sprt']['state']
run['results_stale'] = False
run['results'] = results
self.runs.save(run)
return results
def request_task(self, worker_info):
# Check for blocked user or ip
if self.userdb.is_blocked(worker_info):
return {'task_waiting': False}
max_threads = int(worker_info['concurrency'])
exclusion_list = []
# Does this worker have a task already? If so, just hand that back
existing_run = self.runs.find_one({'tasks': {'$elemMatch': {'active': True, 'worker_info': worker_info}}})
if existing_run != None and existing_run['_id'] not in exclusion_list:
for task_id, task in enumerate(existing_run['tasks']):
if task['active'] and task['worker_info'] == worker_info:
if task['pending']:
return {'run': existing_run, 'task_id': task_id}
else:
# Don't hand back tasks that have been marked as no longer pending
task['active'] = False
self.runs.save(existing_run)
# We need to allocate a new task, but first check we don't have the same
# machine already running because multiple connections are not allowed.
remote_addr = worker_info['remote_addr']
machines = self.get_machines()
connections = sum([int(m.get('remote_addr','') == remote_addr) for m in machines])
# Allow a few connections, for multiple computers on same IP
if connections >= self.userdb.get_machine_limit(worker_info['username']):
return {'task_waiting': False, 'hit_machine_limit': True}
# Ok, we get a new task that does not require more threads than available concurrency
q = {
'new': True,
'query': { '$and': [ {'tasks': {'$elemMatch': {'active': False, 'pending': True}}},
{'args.threads': { '$lte': max_threads }},
{'_id': { '$nin': exclusion_list}},
{'approved': True}]},
'sort': [('args.priority', DESCENDING), ('args.internal_priority', DESCENDING), ('_id', ASCENDING)],
'update': {
'$set': {
'tasks.$.active': True,
'tasks.$.last_updated': datetime.utcnow(),
'tasks.$.worker_info': worker_info,
}
}
}
run = self.runs.find_and_modify(**q)
if run == None:
return {'task_waiting': False}
# Find the task we have just activated: the one with the highest 'last_updated'
latest_time = datetime.min
for idx, task in enumerate(run['tasks']):
if 'last_updated' in task and task['last_updated'] > latest_time:
latest_time = task['last_updated']
task_id = idx
# Recalculate internal priority based on task start date and throughput
# Formula: - second_since_epoch - played_and_allocated_tasks * 3600 * 1000 / games_throughput
# With default value 'throughput = 1000', this means that the priority is unchanged as long as we play at rate '1000 games / hour'.
if (run['args']['throughput'] != None and run['args']['throughput'] != 0):
run['args']['internal_priority'] = - time.mktime(run['start_time'].timetuple()) - task_id * 3600 * 1000 / run['args']['throughput']
self.runs.save(run)
return {'run': run, 'task_id': task_id}
def update_task(self, run_id, task_id, stats, nps, spsa):
run = self.get_run(run_id)
if task_id >= len(run['tasks']):
return {'task_alive': False}
task = run['tasks'][task_id]
if not task['active'] or not task['pending']:
return {'task_alive': False}
# Guard against incorrect results
num_games = stats['wins'] + stats['losses'] + stats['draws']
if 'stats' in task and num_games < task['stats']['wins'] + task['stats']['losses'] + task['stats']['draws']:
return {'task_alive': False}
task['stats'] = stats
task['nps'] = nps
if num_games >= task['num_games']:
task['active'] = False
task['pending'] = False
update_time = datetime.utcnow()
task['last_updated'] = update_time
run['last_updated'] = update_time
run['results_stale'] = True
# Update spsa results
if 'spsa' in run['args'] and spsa['wins'] + spsa['losses'] + spsa['draws'] == spsa['num_games']:
self.update_spsa(run, spsa)
self.runs.save(run)
# Check if SPRT stopping is enabled
if 'sprt' in run['args']:
sprt = run['args']['sprt']
sprt_stats = stat_util.SPRT(self.get_results(run),
elo0=sprt['elo0'],
alpha=sprt['alpha'],
elo1=sprt['elo1'],
beta=sprt['beta'],
drawelo=sprt['drawelo'])
if sprt_stats['finished']:
run['args']['sprt']['state'] = sprt_stats['state']
self.runs.save(run)
self.stop_run(run_id)
return {'task_alive': task['active']}
def failed_task(self, run_id, task_id):
run = self.get_run(run_id)
if task_id >= len(run['tasks']):
return {'task_alive': False}
task = run['tasks'][task_id]
if not task['active'] or not task['pending']:
return {'task_alive': False}
# Mark the task as inactive: it will be rescheduled
task['active'] = False
self.runs.save(run)
return {}
def stop_run(self, run_id):
run = self.get_run(run_id)
prune_idx = len(run['tasks'])
for idx, task in enumerate(run['tasks']):
is_active = task['active']
task['active'] = False
task['pending'] = False
if 'stats' not in task and not is_active:
prune_idx = min(idx, prune_idx)
else:
prune_idx = idx + 1
# Truncate the empty tasks
if prune_idx < len(run['tasks']):
del run['tasks'][prune_idx:]
self.runs.save(run)
return {}
def approve_run(self, run_id, approver):
run = self.get_run(run_id)
# Can't self approve
if run['args']['username'] == approver:
return False
run['approved'] = True
run['approver'] = approver
self.runs.save(run)
return True
def request_spsa(self, run_id, task_id):
run = self.get_run(run_id)
if task_id >= len(run['tasks']):
return {'task_alive': False}
task = run['tasks'][task_id]
if not task['active'] or not task['pending']:
return {'task_alive': False}
result = {
'task_alive': True,
'w_params': [],
'b_params': [],
}
spsa = run['args']['spsa']
# Increment the iter counter
spsa['iter'] += 1
self.runs.save(run)
# Generate the next set of tuning parameters
for param in spsa['params']:
a = param['a'] / (spsa['A'] + spsa['iter']) ** spsa['alpha']
c = param['c'] / spsa['iter'] ** spsa['gamma']
R = a / c ** 2
flip = 1 if bool(random.getrandbits(1)) else -1
result['w_params'].append({
'name': param['name'],
'value': min(max(param['theta'] + c * flip, param['min']), param['max']),
'R': R,
'c': c,
'flip': flip,
})
result['b_params'].append({
'name': param['name'],
'value': min(max(param['theta'] - c * flip, param['min']), param['max']),
})
return result
def update_spsa(self, run, spsa_results):
spsa = run['args']['spsa']
spsa['iter'] += int(spsa_results['num_games'] / 2) - 1
# Update the current theta based on the results from the worker
# Worker wins/losses are always in terms of w_params
result = spsa_results['wins'] - spsa_results['losses']
summary = []
for idx, param in enumerate(spsa['params']):
R = spsa_results['w_params'][idx]['R']
c = spsa_results['w_params'][idx]['c']
flip = spsa_results['w_params'][idx]['flip']
param['theta'] = min(max(param['theta'] + R * c * result * flip, param['min']), param['max'])
summary.append({
'theta': param['theta'],
'R': R,
'c': c,
})
# Every 100 iterations, record the parameters
if 'param_history' not in spsa:
spsa['param_history'] = []
if len(spsa['param_history']) < spsa['iter'] / 100:
spsa['param_history'].append(summary)
class RunDb:
def __init__(self, db_name='fishtest_new'):
# MongoDB server is assumed to be on the same machine, if not user should use
# ssh with port forwarding to access the remote host.
self.conn = MongoClient(os.getenv('FISHTEST_HOST') or 'localhost')
self.db = self.conn[db_name]
self.userdb = UserDb(self.db)
self.actiondb = ActionDb(self.db)
self.pgndb = self.db['pgns']
self.runs = self.db['runs']
self.old_runs = self.db['old_runs']
self.deltas = self.db['deltas']
self.chunk_size = 250
def build_indices(self):
self.runs.ensure_index([('finished', ASCENDING),
('last_updated', DESCENDING)])
self.pgndb.ensure_index([('run_id', ASCENDING)])
def generate_tasks(self, num_games):
tasks = []
remaining = num_games
while remaining > 0:
task_size = min(self.chunk_size, remaining)
tasks.append({
'num_games': task_size,
'pending': True,
'active': False,
})
remaining -= task_size
return tasks
def new_run(self,
base_tag,
new_tag,
num_games,
tc,
book,
book_depth,
threads,
base_options,
new_options,
info='',
resolved_base='',
resolved_new='',
msg_base='',
msg_new='',
base_signature='',
new_signature='',
start_time=None,
sprt=None,
spsa=None,
username=None,
tests_repo=None,
auto_purge=True,
throughput=3000,
priority=0):
if start_time == None:
start_time = datetime.utcnow()
run_args = {
'base_tag': base_tag,
'new_tag': new_tag,
'num_games': num_games,
'tc': tc,
'book': book,
'book_depth': book_depth,
'threads': threads,
'resolved_base': resolved_base,
'resolved_new': resolved_new,
'msg_base': msg_base,
'msg_new': msg_new,
'base_options': base_options,
'new_options': new_options,
'info': info,
'base_signature': base_signature,
'new_signature': new_signature,
'username': username,
'tests_repo': tests_repo,
'auto_purge': auto_purge,
'throughput': throughput,
'priority': priority,
'internal_priority': -time.mktime(start_time.timetuple()),
}
if sprt != None:
run_args['sprt'] = sprt
if spsa != None:
run_args['spsa'] = spsa
new_run = {
'args': run_args,
'start_time': start_time,
'last_updated': start_time,
# Will be filled in by tasks, indexed by task-id
'tasks': self.generate_tasks(num_games),
# Aggregated results
'results': {
'wins': 0,
'losses': 0,
'draws': 0
},
'results_stale': False,
'finished': False,
'approved': False,
'approver': '',
}
return self.runs.insert(new_run)
def get_machines(self):
machines = []
for run in self.runs.find({'tasks': {'$elemMatch': {'active': True}}}):
for task in run['tasks']:
if task['active']:
machine = copy.copy(task['worker_info'])
machine['last_updated'] = task.get('last_updated', None)
machine['run'] = run
machine['nps'] = task.get('nps', 0)
machines.append(machine)
return machines
def get_pgn(self, id):
id = id.split('.')[0] # strip .pgn
pgn = self.pgndb.find_one({'run_id': id})
if pgn:
return zlib.decompress(pgn['pgn_zip']).decode()
return None
def get_pgn_100(self, skip):
return [
p['run_id'] for p in self.pgndb.find(
skip=skip, limit=100, sort=[('_id', DESCENDING)])
]
# Cache runs
run_cache = {}
run_cache_lock = threading.Lock()
run_cache_write_lock = threading.Lock()
timer = None
def get_run(self, id):
with self.run_cache_lock:
id = str(id)
if id in self.run_cache:
self.run_cache[id]['rtime'] = time.time()
return self.run_cache[id]['run']
run = self.runs.find_one({'_id': ObjectId(id)})
if not run:
run = self.old_runs.find_one({'_id': ObjectId(id)})
self.run_cache[id] = {
'rtime': time.time(),
'ftime': time.time(),
'run': run,
'dirty': False
}
return run
def buffer(self, run, flush):
with self.run_cache_lock:
if self.timer is None:
self.timer = threading.Timer(1.0, self.flush_buffers)
self.timer.start()
id = str(run['_id'])
if flush:
self.run_cache[id] = {
'dirty': False,
'rtime': time.time(),
'ftime': time.time(),
'run': run
}
with self.run_cache_write_lock:
self.runs.save(run)
else:
if id in self.run_cache:
ftime = self.run_cache[id]['ftime']
else:
ftime = time.time()
self.run_cache[id] = {
'dirty': True,
'rtime': time.time(),
'ftime': ftime,
'run': run
}
def stop(self):
with self.run_cache_lock:
self.timer = None
time.sleep(1.1)
def flush_buffers(self):
with self.run_cache_lock:
if self.timer is None:
return
now = time.time()
old = now + 1
oldest = None
for id in self.run_cache.keys():
if not self.run_cache[id]['dirty']:
if self.run_cache[id]['rtime'] < now - 60:
del self.run_cache[id]
elif self.run_cache[id]['ftime'] < old:
old = self.run_cache[id]['ftime']
oldest = id
if not oldest is None:
if int(now) % 60 == 0:
self.scavenge(self.run_cache[oldest]['run'])
with self.run_cache_write_lock:
self.runs.save(self.run_cache[oldest]['run'])
self.run_cache[oldest]['dirty'] = False
self.run_cache[oldest]['ftime'] = time.time()
self.timer = threading.Timer(1.0, self.flush_buffers)
self.timer.start()
def scavenge(self, run):
old = datetime.utcnow() - timedelta(minutes=30)
for task in run['tasks']:
if task['active'] and task['last_updated'] < old:
task['active'] = False
def get_runs(self):
return list(self.get_unfinished_runs()) + self.get_finished_runs()[0]
def get_unfinished_runs(self):
with self.run_cache_write_lock:
return self.runs.find({'finished': False},
sort=[('last_updated', DESCENDING),
('start_time', DESCENDING)])
def get_finished_runs(self,
skip=0,
limit=0,
username='',
success_only=False,
ltc_only=False):
q = {'finished': True, 'deleted': {'$exists': False}}
if len(username) > 0:
q['args.username'] = username
if ltc_only:
q['args.tc'] = {'$regex': '^([4-9][0-9])|([1-9][0-9][0-9])'}
if success_only:
# This is unfortunate, but the only way we have of telling if a run was successful or
# not currently is the color!
q['results_info.style'] = '#44EB44'
c = self.runs.find(q,
skip=skip,
limit=limit,
sort=[('last_updated', DESCENDING)])
result = [list(c), c.count()]
if limit != 0 and len(result[0]) != limit:
c = self.old_runs.find(q,
skip=max(0, skip - c.count()),
limit=limit - len(result[0]))
result[0] += list(c)
result[1] += c.count()
else:
result[1] += self.old_runs.find(q).count()
return result
def get_results(self, run, save_run=True):
if not run['results_stale']:
return run['results']
results = {
'wins': 0,
'losses': 0,
'draws': 0,
'crashes': 0,
'time_losses': 0
}
for task in run['tasks']:
if 'stats' in task:
stats = task['stats']
results['wins'] += stats['wins']
results['losses'] += stats['losses']
results['draws'] += stats['draws']
results['crashes'] += stats['crashes']
results['time_losses'] += stats.get('time_losses', 0)
if 'sprt' in run['args'] and 'state' in run['args']['sprt']:
results['sprt'] = run['args']['sprt']['state']
run['results_stale'] = False
run['results'] = results
if save_run:
self.buffer(run, True)
return results
def recalc_prio(self, run, task_id=None):
if task_id is None:
task_id = -1
for task in run['tasks']:
task_id = task_id + 1
if not task['active'] and task['pending']:
break
# Recalculate internal priority based on task start date and throughput
# Formula: - second_since_epoch - played_and_allocated_tasks * 3600 * chunk_size / games_throughput
# With default value 'throughput = 3000', this means that the priority is unchanged as long as
# we play at rate '3000 games / hour'.
if (run['args']['throughput'] != None
and run['args']['throughput'] != 0):
run['args']['internal_priority'] = - time.mktime(run['start_time'].timetuple()) - \
task_id * 3600 * self.chunk_size * run['args']['threads'] / run['args']['throughput']
# Limit concurrent request_task
task_lock = threading.Lock()
task_semaphore = threading.Semaphore(4)
task_time = 0
task_runs = None
def request_task(self, worker_info):
if self.task_semaphore.acquire(False):
try:
with self.task_lock:
return self.sync_request_task(worker_info)
finally:
self.task_semaphore.release()
else:
return {'task_waiting': False}
def sync_request_task(self, worker_info):
if time.time() > self.task_time + 60:
self.task_runs = []
for r in self.get_unfinished_runs():
self.task_runs.append(r)
self.task_runs.sort(key=lambda r: (-r['args']['priority'], -r[
'args']['internal_priority'], r['_id']))
self.task_time = time.time()
max_threads = int(worker_info['concurrency'])
min_threads = int(worker_info.get('min_threads', 1))
max_memory = int(worker_info.get('max_memory', 0))
exclusion_list = []
# We need to allocate a new task, but first check we don't have the same
# machine already running because multiple connections are not allowed.
connections = 0
for run in self.task_runs:
for task in run['tasks']:
if task['active'] and task['worker_info'][
'remote_addr'] == worker_info['remote_addr']:
connections = connections + 1
# Allow a few connections, for multiple computers on same IP
if connections >= self.userdb.get_machine_limit(
worker_info['username']):
return {'task_waiting': False, 'hit_machine_limit': True}
# Get a new task that matches the worker requirements
run_found = False
for runt in self.task_runs:
run = self.get_run(runt['_id'])
# compute required TT memory
need_tt = 0
if max_memory > 0:
def get_hash(s):
h = re.search('Hash=([0-9]+)', s)
if h:
return int(h.group(1))
return 0
need_tt += get_hash(run['args']['new_options'])
need_tt += get_hash(run['args']['base_options'])
need_tt *= max_threads // run['args']['threads']
if run['_id'] not in exclusion_list and run['approved'] \
and run['args']['threads'] <= max_threads \
and run['args']['threads'] >= min_threads \
and need_tt <= max_memory:
task_id = -1
for task in run['tasks']:
task_id = task_id + 1
if not task['active'] and task['pending']:
task['worker_info'] = worker_info
task['last_updated'] = datetime.utcnow()
task['active'] = True
run_found = True
break
if run_found:
break
if not run_found:
return {'task_waiting': False}
self.recalc_prio(run, task_id)
self.buffer(run, False)
for runt in self.task_runs:
if runt['_id'] == run['_id']:
runt['args']['internal_priority'] = run['args'][
'internal_priority']
self.task_runs.sort(key=lambda r: (-r['args']['priority'], -r[
'args']['internal_priority'], r['_id']))
break
return {'run': run, 'task_id': task_id}
# Create a lock for each active run
run_lock = threading.Lock()
active_runs = {}
purge_count = 0
def active_run_lock(self, id):
with self.run_lock:
self.purge_count = self.purge_count + 1
if self.purge_count > 100000:
old = time.time() - 10000
self.active_runs = dict(
(k, v) for k, v in self.active_runs.iteritems()
if v['time'] >= old)
self.purge_count = 0
if id in self.active_runs:
active_lock = self.active_runs[id]['lock']
self.active_runs[id]['time'] = time.time()
else:
active_lock = threading.Lock()
self.active_runs[id] = {
'time': time.time(),
'lock': active_lock
}
return active_lock
def update_task(self, run_id, task_id, stats, nps, spsa, username):
lock = self.active_run_lock(str(run_id))
with lock:
return self.sync_update_task(run_id, task_id, stats, nps, spsa,
username)
def sync_update_task(self, run_id, task_id, stats, nps, spsa, username):
run = self.get_run(run_id)
if task_id >= len(run['tasks']):
return {'task_alive': False}
task = run['tasks'][task_id]
if not task['active'] or not task['pending']:
return {'task_alive': False}
if task['worker_info']['username'] != username:
print('Update_task: Non matching username: '******'task_alive': False}
# Guard against incorrect results
count_games = lambda d: d['wins'] + d['losses'] + d['draws']
num_games = count_games(stats)
old_num_games = count_games(
task['stats']) if 'stats' in task else num_games
spsa_games = count_games(spsa) if 'spsa' in run['args'] else 0
if num_games < old_num_games \
or (spsa_games > 0 and num_games <= 0) \
or (spsa_games > 0 and 'stats' in task and num_games <= old_num_games):
return {'task_alive': False}
flush = False
task['stats'] = stats
task['nps'] = nps
if num_games >= task['num_games']:
task[
'pending'] = False # Make pending False before making active false to prevent race in request_task
task['active'] = False
flush = True
update_time = datetime.utcnow()
task['last_updated'] = update_time
run['last_updated'] = update_time
run['results_stale'] = True
# Update spsa results
if 'spsa' in run['args'] and spsa_games == spsa['num_games']:
self.update_spsa(task['worker_info']['unique_key'], run, spsa)
# Check if SPRT stopping is enabled
if 'sprt' in run['args']:
sprt = run['args']['sprt']
sprt_stats = fishtest.stat_util.SPRT(self.get_results(run, False),
elo0=sprt['elo0'],
alpha=sprt['alpha'],
elo1=sprt['elo1'],
beta=sprt['beta'],
drawelo=sprt['drawelo'])
if sprt_stats['finished']:
run['args']['sprt']['state'] = sprt_stats['state']
self.stop_run(run_id, run)
flush = True
self.buffer(run, flush)
return {'task_alive': task['active']}
def upload_pgn(self, run_id, pgn_zip):
self.pgndb.insert({'run_id': run_id, 'pgn_zip': Binary(pgn_zip)})
return {}
def failed_task(self, run_id, task_id):
run = self.get_run(run_id)
if task_id >= len(run['tasks']):
return {'task_alive': False}
task = run['tasks'][task_id]
if not task['active'] or not task['pending']:
return {'task_alive': False}
# Mark the task as inactive: it will be rescheduled
task['active'] = False
self.buffer(run, True)
return {}
def stop_run(self, run_id, run=None):
self.clear_params(run_id)
save_it = False
if run is None:
run = self.get_run(run_id)
save_it = True
prune_idx = len(run['tasks'])
for idx, task in enumerate(run['tasks']):
is_active = task['active']
task[
'pending'] = False # Make pending False before making active false to prevent race in request_task
task['active'] = False
if 'stats' not in task and not is_active:
prune_idx = min(idx, prune_idx)
else:
prune_idx = idx + 1
# Truncate the empty tasks
if prune_idx < len(run['tasks']):
del run['tasks'][prune_idx:]
if save_it:
self.buffer(run, True)
self.task_time = 0
return {}
def approve_run(self, run_id, approver):
run = self.get_run(run_id)
# Can't self approve
if run['args']['username'] == approver:
return False
run['approved'] = True
run['approver'] = approver
self.buffer(run, True)
self.task_time = 0
return True
def spsa_param_clip_round(self, param, increment, clipping, rounding):
value = 0.0
if clipping == 'old':
value = min(max(param['theta'] + increment, param['min']),
param['max'])
else: #clipping == 'careful':
inc = min(abs(increment),
abs(param['theta'] - param['min']) / 2,
abs(param['theta'] - param['max']) / 2)
if inc > 0:
inc_sgn = 0 if increment == 0 else increment / abs(increment)
value = param['theta'] + inc_sgn * inc
else: #revert to old behavior to bounce off boundary
value = min(max(param['theta'] + increment, param['min']),
param['max'])
#'deterministic' rounding calls round() inside the worker.
#'randomized' says 4.p should be 5 with probability p, 4 with probability 1-p,
# and is continuous (albeit after expectation) unlike round().
if rounding == 'randomized':
fl = math.floor(
value) #greatest integer <= value, thus works for negative.
if random.uniform(0, 1) < value - fl:
value = fl + 1
else:
value = fl
return value
# Store SPSA parameters for each worker
spsa_params = {}
def store_params(self, run_id, worker, params):
run_id = str(run_id)
if not run_id in self.spsa_params:
self.spsa_params[run_id] = {}
self.spsa_params[run_id][worker] = params
def get_params(self, run_id, worker):
run_id = str(run_id)
if not run_id in self.spsa_params:
# Should only happen after server restart
return self.generate_spsa(self.get_run(run_id))['w_params']
return self.spsa_params[run_id][worker]
def clear_params(self, run_id):
run_id = str(run_id)
if run_id in self.spsa_params:
del self.spsa_params[run_id]
def request_spsa(self, run_id, task_id):
run = self.get_run(run_id)
if task_id >= len(run['tasks']):
return {'task_alive': False}
task = run['tasks'][task_id]
if not task['active'] or not task['pending']:
return {'task_alive': False}
result = self.generate_spsa(run)
self.store_params(run['_id'], task['worker_info']['unique_key'],
result['w_params'])
return result
def generate_spsa(self, run):
result = {
'task_alive': True,
'w_params': [],
'b_params': [],
}
spsa = run['args']['spsa']
if 'clipping' not in spsa:
spsa['clipping'] = 'old'
if 'rounding' not in spsa:
spsa['rounding'] = 'deterministic'
# Generate the next set of tuning parameters
iter_local = spsa[
'iter'] + 1 #assume at least one completed, and avoid division by zero
for param in spsa['params']:
a = param['a'] / (spsa['A'] + iter_local)**spsa['alpha']
c = param['c'] / iter_local**spsa['gamma']
R = a / c**2
flip = 1 if bool(random.getrandbits(1)) else -1
result['w_params'].append({
'name':
param['name'],
'value':
self.spsa_param_clip_round(param, c * flip, spsa['clipping'],
spsa['rounding']),
'R':
R,
'c':
c,
'flip':
flip,
})
result['b_params'].append({
'name':
param['name'],
'value':
self.spsa_param_clip_round(param, -c * flip, spsa['clipping'],
spsa['rounding']),
})
return result
def update_spsa(self, worker, run, spsa_results):
spsa = run['args']['spsa']
if 'clipping' not in spsa:
spsa['clipping'] = 'old'
spsa['iter'] += int(spsa_results['num_games'] / 2)
# Store the history every 'freq' iterations.
# More tuned parameters result in a lower update frequency,
# so that the required storage (performance) remains constant.
if 'param_history' not in spsa:
spsa['param_history'] = []
L = len(spsa['params'])
freq = L * 25
if freq < 100:
freq = 100
maxlen = 250000 / freq
grow_summary = len(spsa['param_history']) < min(
maxlen, spsa['iter'] / freq)
# Update the current theta based on the results from the worker
# Worker wins/losses are always in terms of w_params
result = spsa_results['wins'] - spsa_results['losses']
summary = []
w_params = self.get_params(run['_id'], worker)
for idx, param in enumerate(spsa['params']):
R = w_params[idx]['R']
c = w_params[idx]['c']
flip = w_params[idx]['flip']
param['theta'] = self.spsa_param_clip_round(
param, R * c * result * flip, spsa['clipping'],
'deterministic')
if grow_summary:
summary.append({
'theta': param['theta'],
'R': R,
'c': c,
})
if grow_summary:
spsa['param_history'].append(summary)
class RunDb:
def __init__(self, db_name='fishtest_new', clop_socket=None):
# MongoDB server is assumed to be on the same machine, if not user should use
# ssh with port forwarding to access the remote host.
self.conn = MongoClient(os.getenv('FISHTEST_HOST') or 'localhost')
self.db = self.conn[db_name]
self.userdb = UserDb(self.db)
self.clopdb = ClopDb(self.db, clop_socket)
self.actiondb = ActionDb(self.db)
self.runs = self.db['runs']
self.chunk_size = 1000
def build_indices(self):
self.runs.ensure_index([('finished', ASCENDING), ('last_updated', DESCENDING)])
def generate_tasks(self, num_games):
tasks = []
remaining = num_games
while remaining > 0:
task_size = min(self.chunk_size, remaining)
tasks.append({
'num_games': task_size,
'pending': True,
'active': False,
})
remaining -= task_size
return tasks
def new_run(self, base_tag, new_tag, num_games, tc, book, book_depth, threads, base_options, new_options,
info='',
resolved_base='',
resolved_new='',
msg_base='',
msg_new='',
base_signature='',
new_signature='',
regression_test=False,
start_time=None,
sprt=None,
clop=None,
username=None,
tests_repo=None,
priority=0):
if start_time == None:
start_time = datetime.utcnow()
run_args = {
'base_tag': base_tag,
'new_tag': new_tag,
'num_games': num_games,
'tc': tc,
'book': book,
'book_depth': book_depth,
'threads': threads,
'regression_test': regression_test,
'resolved_base': resolved_base,
'resolved_new': resolved_new,
'msg_base': msg_base,
'msg_new': msg_new,
'base_options': base_options,
'new_options': new_options,
'info': info,
'base_signature': base_signature,
'new_signature': new_signature,
'username': username,
'tests_repo': tests_repo,
'priority': priority,
}
if sprt != None:
run_args['sprt'] = sprt
if clop != None:
run_args['clop'] = clop
new_run = {
'args': run_args,
'start_time': start_time,
'last_updated': start_time,
# Will be filled in by tasks, indexed by task-id
'tasks': self.generate_tasks(num_games),
# Aggregated results
'results': { 'wins': 0, 'losses': 0, 'draws': 0 },
'results_stale': False,
'finished': False,
'approved': False,
'approver': '',
}
# Check for an existing approval matching the git commit SHAs
def get_approval(sha):
q = { '$or': [{ 'args.resolved_base': sha }, { 'args.resolved_new': sha }], 'approved': True }
return self.runs.find_one(q)
base_approval = get_approval(resolved_base)
new_approval = get_approval(resolved_new)
if base_approval != None and new_approval != None:
new_run['approved'] = True
new_run['approver'] = new_approval['approver']
return self.runs.insert(new_run)
def get_machines(self):
machines = []
for run in self.runs.find({'tasks': {'$elemMatch': {'active': True}}}):
for task in run['tasks']:
if task['active']:
machine = copy.copy(task['worker_info'])
machine['last_updated'] = task.get('last_updated', None)
machine['run'] = run
machine['nps'] = task.get('nps', 0)
# TODO(glinscott): Temporary - remove once worker version >= 41
if not isinstance(machine['uname'], basestring):
machine['uname'] = machine['uname'][0] + machine['uname'][2]
machines.append(machine)
return machines
def get_run(self, id):
return self.runs.find_one({'_id': ObjectId(id)})
def get_run_to_build(self):
return self.runs.find_one({'binaries_url': {'$exists': False}, 'finished': False, 'deleted': {'$exists': False}})
def get_runs(self):
return list(self.get_unfinished_runs()) + self.get_finished_runs()[0]
def get_unfinished_runs(self):
return self.runs.find({'finished': False},
sort=[('last_updated', DESCENDING), ('start_time', DESCENDING)])
def get_finished_runs(self, skip=0, limit=0, username=''):
q = {'finished': True, 'deleted': {'$exists': False}}
if len(username) > 0:
q['args.username'] = username
c = self.runs.find(q, skip=skip, limit=limit, sort=[('last_updated', DESCENDING)])
return (list(c), c.count())
def get_clop_exclusion_list(self, minimum):
exclusion_list = []
for run in self.runs.find({'args.clop': {'$exists': True}, 'finished': False, 'deleted': {'$exists': False}}):
available_games = 0
for game in self.clopdb.get_games(run['_id']):
if len(game['task_id']) == 0:
available_games += 1
active = False
for task in run['tasks']:
if task['active']:
active = True
if available_games < minimum and active:
exclusion_list.append(run['_id'])
return exclusion_list
def get_results(self, run):
if not run['results_stale']:
return run['results']
results = { 'wins': 0, 'losses': 0, 'draws': 0, 'crashes': 0, 'time_losses':0 }
for task in run['tasks']:
if 'stats' in task:
stats = task['stats']
results['wins'] += stats['wins']
results['losses'] += stats['losses']
results['draws'] += stats['draws']
results['crashes'] += stats['crashes']
results['time_losses'] += stats.get('time_losses', 0)
if 'sprt' in run['args'] and 'state' in run['args']['sprt']:
results['sprt'] = run['args']['sprt']['state']
run['results_stale'] = False
run['results'] = results
self.runs.save(run)
return results
def request_task(self, worker_info):
# Check for blocked user or ip
if self.userdb.is_blocked(worker_info):
return {'task_waiting': False}
# Build list of CLOP runs that are already almost full
max_threads = int(worker_info['concurrency'])
if 'Windows' in worker_info['uname']:
exclusion_list = [r['_id'] for r in self.runs.find({'args.clop': {'$exists': True}, 'finished': False, 'deleted': {'$exists': False}})]
else:
exclusion_list = self.get_clop_exclusion_list(2 + max_threads)
# Does this worker have a task already? If so, just hand that back
existing_run = self.runs.find_one({'tasks': {'$elemMatch': {'active': True, 'worker_info': worker_info}}})
if existing_run != None and existing_run['_id'] not in exclusion_list:
for task_id, task in enumerate(existing_run['tasks']):
if task['active'] and task['worker_info'] == worker_info:
if task['pending']:
return {'run': existing_run, 'task_id': task_id}
else:
# Don't hand back tasks that have been marked as no longer pending
task['active'] = False
self.runs.save(existing_run)
# We need to allocate a new task, but first check we don't have the same
# machine already running because multiple connections are not allowed.
remote_addr = worker_info['remote_addr']
machines = self.get_machines()
connections = sum([int(m.get('remote_addr','') == remote_addr) for m in machines])
# Allow a few connections, for multiple computers on same IP
if connections >= self.userdb.get_machine_limit(worker_info['username']):
return {'task_waiting': False, 'hit_machine_limit': True}
# Ok, we get a new task that does not require more threads than available concurrency
q = {
'new': True,
'query': { '$and': [ {'tasks': {'$elemMatch': {'active': False, 'pending': True}}},
{'args.threads': { '$lte': max_threads }},
{'_id': { '$nin': exclusion_list}},
{'approved': True}]},
'sort': [('args.priority', DESCENDING), ('_id', ASCENDING)],
'update': {
'$set': {
'tasks.$.active': True,
'tasks.$.last_updated': datetime.utcnow(),
'tasks.$.worker_info': worker_info,
}
}
}
run = self.runs.find_and_modify(**q)
if run == None:
return {'task_waiting': False}
# Find the task we have just activated: the one with the highest 'last_updated'
latest_time = datetime.min
for idx, task in enumerate(run['tasks']):
if 'last_updated' in task and task['last_updated'] > latest_time:
latest_time = task['last_updated']
task_id = idx
# Lower priority of long running tests
if task_id > 40 and 'sprt' in run['args'] and run['args']['priority'] == 0:
run['args']['priority'] = -1
self.runs.save(run)
# Lower priority of LTC long running tests
if task_id > 50 and 'sprt' in run['args'] and parse_tc(run['args']['tc']) > 100:
run['args']['priority'] = -3
self.runs.save(run)
return {'run': run, 'task_id': task_id}
def update_task(self, run_id, task_id, stats, nps, clop):
run = self.get_run(run_id)
if task_id >= len(run['tasks']):
return {'task_alive': False}
task = run['tasks'][task_id]
if not task['active'] or not task['pending']:
return {'task_alive': False}
# Guard against incorrect results
num_games = stats['wins'] + stats['losses'] + stats['draws']
if 'stats' in task and num_games < task['stats']['wins'] + task['stats']['losses'] + task['stats']['draws']:
return {'task_alive': False}
task['stats'] = stats
task['nps'] = nps
if num_games >= task['num_games']:
task['active'] = False
task['pending'] = False
update_time = datetime.utcnow()
task['last_updated'] = update_time
run['last_updated'] = update_time
run['results_stale'] = True
self.runs.save(run)
# Check if SPRT stopping is enabled
if 'sprt' in run['args']:
sprt = run['args']['sprt']
sprt_stats = stat_util.SPRT(self.get_results(run),
elo0=sprt['elo0'],
alpha=sprt['alpha'],
elo1=sprt['elo1'],
beta=sprt['beta'],
drawelo=sprt['drawelo'])
if sprt_stats['finished']:
run['args']['sprt']['state'] = sprt_stats['state']
self.runs.save(run)
self.stop_run(run_id)
# Update clop results
if 'clop' in run['args'] and len(clop['game_id']) > 0:
self.clopdb.write_result(clop['game_id'], clop['game_result'])
if not task['active']:
self.clopdb.stop_games(run_id, task_id)
return {'task_alive': task['active']}
def failed_task(self, run_id, task_id):
run = self.get_run(run_id)
if task_id >= len(run['tasks']):
return {'task_alive': False}
task = run['tasks'][task_id]
if not task['active'] or not task['pending']:
return {'task_alive': False}
# Mark the task as inactive: it will be rescheduled
task['active'] = False
self.runs.save(run)
if 'clop' in run['args']:
self.clopdb.stop_games(run_id, task_id)
return {}
def stop_run(self, run_id):
run = self.get_run(run_id)
prune_idx = len(run['tasks'])
for idx, task in enumerate(run['tasks']):
is_active = task['active']
task['active'] = False
task['pending'] = False
if 'stats' not in task and not is_active:
prune_idx = min(idx, prune_idx)
else:
prune_idx = idx + 1
# Truncate the empty tasks
if prune_idx < len(run['tasks']):
del run['tasks'][prune_idx:]
self.runs.save(run)
return {}
def approve_run(self, run_id, approver):
run = self.get_run(run_id)
# Can't self approve
if run['args']['username'] == approver:
return False
run['approved'] = True
run['approver'] = approver
self.runs.save(run)
return True
class RunDb:
def __init__(self, db_name='fishtest_new'):
# MongoDB server is assumed to be on the same machine, if not user should use
# ssh with port forwarding to access the remote host.
self.conn = MongoClient(os.getenv('FISHTEST_HOST') or 'localhost')
self.db = self.conn[db_name]
self.userdb = UserDb(self.db)
self.actiondb = ActionDb(self.db)
self.pgndb = self.db['pgns']
self.runs = self.db['runs']
self.old_runs = self.db['old_runs']
self.chunk_size = 250
def build_indices(self):
self.runs.ensure_index([('finished', ASCENDING), ('last_updated', DESCENDING)])
self.pgndb.ensure_index([('run_id', ASCENDING)])
def generate_tasks(self, num_games):
tasks = []
remaining = num_games
while remaining > 0:
task_size = min(self.chunk_size, remaining)
tasks.append({
'num_games': task_size,
'pending': True,
'active': False,
})
remaining -= task_size
return tasks
def new_run(self, base_tag, new_tag, num_games, tc, book, book_depth, threads, base_options, new_options,
info='',
resolved_base='',
resolved_new='',
msg_base='',
msg_new='',
base_signature='',
new_signature='',
start_time=None,
sprt=None,
spsa=None,
username=None,
tests_repo=None,
auto_purge=True,
throughput=3000,
priority=0):
if start_time == None:
start_time = datetime.utcnow()
run_args = {
'base_tag': base_tag,
'new_tag': new_tag,
'num_games': num_games,
'tc': tc,
'book': book,
'book_depth': book_depth,
'threads': threads,
'resolved_base': resolved_base,
'resolved_new': resolved_new,
'msg_base': msg_base,
'msg_new': msg_new,
'base_options': base_options,
'new_options': new_options,
'info': info,
'base_signature': base_signature,
'new_signature': new_signature,
'username': username,
'tests_repo': tests_repo,
'auto_purge': auto_purge,
'throughput': throughput,
'priority': priority,
'internal_priority': - time.mktime(start_time.timetuple()),
}
if sprt != None:
run_args['sprt'] = sprt
if spsa != None:
run_args['spsa'] = spsa
new_run = {
'args': run_args,
'start_time': start_time,
'last_updated': start_time,
# Will be filled in by tasks, indexed by task-id
'tasks': self.generate_tasks(num_games),
# Aggregated results
'results': { 'wins': 0, 'losses': 0, 'draws': 0 },
'results_stale': False,
'finished': False,
'approved': False,
'approver': '',
}
return self.runs.insert(new_run)
def get_machines(self):
machines = []
for run in self.runs.find({'tasks': {'$elemMatch': {'active': True}}}):
for task in run['tasks']:
if task['active']:
machine = copy.copy(task['worker_info'])
machine['last_updated'] = task.get('last_updated', None)
machine['run'] = run
machine['nps'] = task.get('nps', 0)
machines.append(machine)
return machines
def get_pgn(self, id):
id = id.split('.')[0] # strip .pgn
pgn = self.pgndb.find_one({'run_id': id})
if pgn:
return zlib.decompress(pgn['pgn_zip']).decode()
return None
def get_pgn_100(self, skip):
return [p['run_id'] for p in self.pgndb.find(skip=skip, limit=100, sort=[('_id',DESCENDING)])]
# Cache runs
run_cache = {}
run_cache_lock = threading.Lock()
run_cache_write_lock = threading.Lock()
timer = None
def get_run(self, id):
with self.run_cache_lock:
id = str(id)
if id in self.run_cache:
self.run_cache[id]['rtime'] = time.time()
return self.run_cache[id]['run']
run = self.runs.find_one({'_id': ObjectId(id)})
if not run:
run = self.old_runs.find_one({'_id': ObjectId(id)})
self.run_cache[id] = { 'rtime': time.time(), 'ftime': time.time(), 'run': run, 'dirty': False }
return run
def buffer(self, run, flush):
with self.run_cache_lock:
if self.timer is None:
self.timer = threading.Timer(1.0, self.flush_buffers)
self.timer.start()
id = str(run['_id'])
if flush:
self.run_cache[id] = { 'dirty': False, 'rtime': time.time(), 'ftime': time.time(), 'run': run }
with self.run_cache_write_lock:
self.runs.save(run)
else:
if id in self.run_cache:
ftime = self.run_cache[id]['ftime']
else:
ftime = time.time()
self.run_cache[id] = { 'dirty': True, 'rtime': time.time(), 'ftime': ftime, 'run': run }
def stop(self):
with self.run_cache_lock:
self.timer = None
time.sleep(1.1)
def flush_buffers(self):
with self.run_cache_lock:
if self.timer is None:
return
now = time.time()
old = now + 1
oldest = None
for id in self.run_cache.keys():
if not self.run_cache[id]['dirty']:
if self.run_cache[id]['rtime'] < now - 60:
del self.run_cache[id]
elif self.run_cache[id]['ftime'] < old:
old = self.run_cache[id]['ftime']
oldest = id
if not oldest is None:
if int(now) % 60 == 0:
self.scavenge(self.run_cache[oldest]['run'])
with self.run_cache_write_lock:
self.runs.save(self.run_cache[oldest]['run'])
self.run_cache[oldest]['dirty'] = False
self.run_cache[oldest]['ftime'] = time.time()
self.timer = threading.Timer(1.0, self.flush_buffers)
self.timer.start()
def scavenge(self, run):
old = datetime.utcnow() - timedelta(minutes=30)
for task in run['tasks']:
if task['active'] and task['last_updated'] < old:
task['active'] = False
def get_runs(self):
return list(self.get_unfinished_runs()) + self.get_finished_runs()[0]
def get_unfinished_runs(self):
with self.run_cache_write_lock:
return self.runs.find({'finished': False},
sort=[('last_updated', DESCENDING), ('start_time', DESCENDING)])
def get_finished_runs(self, skip=0, limit=0, username='', success_only=False, ltc_only=False):
q = {'finished': True, 'deleted': {'$exists': False}}
if len(username) > 0:
q['args.username'] = username
if ltc_only:
q['args.tc'] = {'$regex':'^([4-9][0-9])|([1-9][0-9][0-9])'}
if success_only:
# This is unfortunate, but the only way we have of telling if a run was successful or
# not currently is the color!
q['results_info.style'] = '#44EB44'
c = self.runs.find(q, skip=skip, limit=limit, sort=[('last_updated', DESCENDING)])
result = [list(c), c.count()]
if limit != 0 and len(result[0]) != limit:
c = self.old_runs.find(q, skip=max(0, skip-c.count()), limit=limit-len(result[0]))
result[0] += list(c)
result[1] += c.count()
else:
result[1] += self.old_runs.find(q).count()
return result
def get_results(self, run, save_run=True):
if not run['results_stale']:
return run['results']
results = { 'wins': 0, 'losses': 0, 'draws': 0, 'crashes': 0, 'time_losses':0 }
for task in run['tasks']:
if 'stats' in task:
stats = task['stats']
results['wins'] += stats['wins']
results['losses'] += stats['losses']
results['draws'] += stats['draws']
results['crashes'] += stats['crashes']
results['time_losses'] += stats.get('time_losses', 0)
if 'sprt' in run['args'] and 'state' in run['args']['sprt']:
results['sprt'] = run['args']['sprt']['state']
run['results_stale'] = False
run['results'] = results
if save_run:
self.buffer(run, True)
return results
def recalc_prio(self, run, task_id=None):
if task_id is None:
task_id = -1
for task in run['tasks']:
task_id = task_id + 1
if not task['active'] and task['pending']:
break
# Recalculate internal priority based on task start date and throughput
# Formula: - second_since_epoch - played_and_allocated_tasks * 3600 * chunk_size / games_throughput
# With default value 'throughput = 3000', this means that the priority is unchanged as long as
# we play at rate '3000 games / hour'.
if (run['args']['throughput'] != None and run['args']['throughput'] != 0):
run['args']['internal_priority'] = - time.mktime(run['start_time'].timetuple()) - \
task_id * 3600 * self.chunk_size * run['args']['threads'] / run['args']['throughput']
# Limit concurrent request_task
task_lock = threading.Lock()
task_semaphore = threading.Semaphore(4)
task_time = 0
task_runs = None
def request_task(self, worker_info):
if self.task_semaphore.acquire(False):
try:
with self.task_lock:
return self.sync_request_task(worker_info)
finally:
self.task_semaphore.release()
else:
return {'task_waiting': False}
def sync_request_task(self, worker_info):
if time.time() > self.task_time + 60:
self.task_runs = []
for r in self.get_unfinished_runs():
self.task_runs.append(r)
self.task_runs.sort(key=lambda r: (-r['args']['priority'], -r['args']['internal_priority'], r['_id']))
self.task_time = time.time()
max_threads = int(worker_info['concurrency'])
min_threads = int(worker_info.get('min_threads', 1))
max_memory = int(worker_info.get('max_memory', 0))
exclusion_list = []
# We need to allocate a new task, but first check we don't have the same
# machine already running because multiple connections are not allowed.
connections = 0
for run in self.task_runs:
for task in run['tasks']:
if task['active'] and task['worker_info']['remote_addr'] == worker_info['remote_addr']:
connections = connections + 1
# Allow a few connections, for multiple computers on same IP
if connections >= self.userdb.get_machine_limit(worker_info['username']):
return {'task_waiting': False, 'hit_machine_limit': True}
# Get a new task that matches the worker requirements
run_found = False
for runt in self.task_runs:
run = self.get_run(runt['_id'])
# compute required TT memory
need_tt = 0
if max_memory > 0:
def get_hash(s):
h = re.search('Hash=([0-9]+)', s)
if h:
return int(h.group(1))
return 0
need_tt += get_hash(run['args']['new_options'])
need_tt += get_hash(run['args']['base_options'])
need_tt *= max_threads // run['args']['threads']
if run['_id'] not in exclusion_list and run['approved'] \
and run['args']['threads'] <= max_threads \
and run['args']['threads'] >= min_threads \
and need_tt <= max_memory:
task_id = -1
for task in run['tasks']:
task_id = task_id + 1
if not task['active'] and task['pending']:
task['worker_info'] = worker_info
task['last_updated'] = datetime.utcnow()
task['active'] = True
run_found = True
break
if run_found:
break
if not run_found:
return {'task_waiting': False}
self.recalc_prio(run, task_id)
self.buffer(run, False)
for runt in self.task_runs:
if runt['_id'] == run['_id']:
runt['args']['internal_priority'] = run['args']['internal_priority']
self.task_runs.sort(key=lambda r: (-r['args']['priority'], -r['args']['internal_priority'], r['_id']))
break
return {'run': run, 'task_id': task_id}
# Create a lock for each active run
run_lock = threading.Lock()
active_runs = {}
purge_count = 0
def active_run_lock(self, id):
with self.run_lock:
self.purge_count = self.purge_count + 1
if self.purge_count > 100000:
old = time.time() - 10000
self.active_runs = dict((k,v) for k, v in self.active_runs.iteritems() if v['time'] >= old)
self.purge_count = 0
if id in self.active_runs:
active_lock = self.active_runs[id]['lock']
self.active_runs[id]['time'] = time.time()
else:
active_lock = threading.Lock()
self.active_runs[id] = { 'time': time.time(), 'lock': active_lock }
return active_lock
def update_task(self, run_id, task_id, stats, nps, spsa, username):
lock = self.active_run_lock(str(run_id))
with lock:
return self.sync_update_task(run_id, task_id, stats, nps, spsa, username)
def sync_update_task(self, run_id, task_id, stats, nps, spsa, username):
run = self.get_run(run_id)
if task_id >= len(run['tasks']):
return {'task_alive': False}
task = run['tasks'][task_id]
if not task['active'] or not task['pending']:
return {'task_alive': False}
if task['worker_info']['username'] != username:
print('Update_task: Non matching username: '******'task_alive': False}
# Guard against incorrect results
count_games = lambda d: d['wins'] + d['losses'] + d['draws']
num_games = count_games(stats)
old_num_games = count_games(task['stats']) if 'stats' in task else num_games
spsa_games = count_games(spsa) if 'spsa' in run['args'] else 0
if num_games < old_num_games \
or (spsa_games > 0 and num_games <= 0) \
or (spsa_games > 0 and 'stats' in task and num_games <= old_num_games):
return {'task_alive': False}
flush = False
task['stats'] = stats
task['nps'] = nps
if num_games >= task['num_games']:
task['pending'] = False # Make pending False before making active false to prevent race in request_task
task['active'] = False
flush = True
update_time = datetime.utcnow()
task['last_updated'] = update_time
run['last_updated'] = update_time
run['results_stale'] = True
# Update spsa results
if 'spsa' in run['args'] and spsa_games == spsa['num_games']:
self.update_spsa(task['worker_info']['unique_key'], run, spsa)
# Check if SPRT stopping is enabled
if 'sprt' in run['args']:
sprt = run['args']['sprt']
sprt_stats = fishtest.stat_util.SPRT(self.get_results(run, False),
elo0=sprt['elo0'],
alpha=sprt['alpha'],
elo1=sprt['elo1'],
beta=sprt['beta'],
drawelo=sprt['drawelo'])
if sprt_stats['finished']:
run['args']['sprt']['state'] = sprt_stats['state']
self.stop_run(run_id, run)
flush = True
self.buffer(run, flush)
return {'task_alive': task['active']}
def upload_pgn(self, run_id, pgn_zip):
self.pgndb.insert({'run_id': run_id, 'pgn_zip': Binary(pgn_zip)})
return {}
def failed_task(self, run_id, task_id):
run = self.get_run(run_id)
if task_id >= len(run['tasks']):
return {'task_alive': False}
task = run['tasks'][task_id]
if not task['active'] or not task['pending']:
return {'task_alive': False}
# Mark the task as inactive: it will be rescheduled
task['active'] = False
self.buffer(run, True)
return {}
def stop_run(self, run_id, run=None):
self.clear_params(run_id)
save_it = False
if run is None:
run = self.get_run(run_id)
save_it = True
prune_idx = len(run['tasks'])
for idx, task in enumerate(run['tasks']):
is_active = task['active']
task['pending'] = False # Make pending False before making active false to prevent race in request_task
task['active'] = False
if 'stats' not in task and not is_active:
prune_idx = min(idx, prune_idx)
else:
prune_idx = idx + 1
# Truncate the empty tasks
if prune_idx < len(run['tasks']):
del run['tasks'][prune_idx:]
if save_it:
self.buffer(run, True)
self.task_time = 0
return {}
def approve_run(self, run_id, approver):
run = self.get_run(run_id)
# Can't self approve
if run['args']['username'] == approver:
return False
run['approved'] = True
run['approver'] = approver
self.buffer(run, True)
self.task_time = 0
return True
def spsa_param_clip_round(self, param, increment, clipping, rounding):
value = 0.0
if clipping == 'old':
value = min(max(param['theta'] + increment, param['min']), param['max'])
else: #clipping == 'careful':
inc = min(abs(increment), abs(param['theta'] - param['min']) / 2, abs(param['theta'] - param['max']) / 2)
if inc > 0:
inc_sgn = 0 if increment == 0 else increment / abs(increment)
value = param['theta'] + inc_sgn * inc
else: #revert to old behavior to bounce off boundary
value = min(max(param['theta'] + increment, param['min']), param['max'])
#'deterministic' rounding calls round() inside the worker.
#'randomized' says 4.p should be 5 with probability p, 4 with probability 1-p,
# and is continuous (albeit after expectation) unlike round().
if rounding == 'randomized':
fl = math.floor(value) #greatest integer <= value, thus works for negative.
if random.uniform(0,1) < value - fl:
value = fl + 1
else:
value = fl
return value
# Store SPSA parameters for each worker
spsa_params = {}
def store_params(self, run_id, worker, params):
run_id = str(run_id)
if not run_id in self.spsa_params:
self.spsa_params[run_id] = {}
self.spsa_params[run_id][worker] = params
def get_params(self, run_id, worker):
run_id = str(run_id)
if not run_id in self.spsa_params:
# Should only happen after server restart
return self.generate_spsa(self.get_run(run_id))['w_params']
return self.spsa_params[run_id][worker]
def clear_params(self, run_id):
run_id = str(run_id)
if run_id in self.spsa_params:
del self.spsa_params[run_id]
def request_spsa(self, run_id, task_id):
run = self.get_run(run_id)
if task_id >= len(run['tasks']):
return {'task_alive': False}
task = run['tasks'][task_id]
if not task['active'] or not task['pending']:
return {'task_alive': False}
result = self.generate_spsa(run)
self.store_params(run['_id'], task['worker_info']['unique_key'], result['w_params'])
return result
def generate_spsa(self, run):
result = {
'task_alive': True,
'w_params': [],
'b_params': [],
}
spsa = run['args']['spsa']
if 'clipping' not in spsa:
spsa['clipping'] = 'old'
if 'rounding' not in spsa:
spsa['rounding'] = 'deterministic'
# Generate the next set of tuning parameters
iter_local = spsa['iter'] + 1 #assume at least one completed, and avoid division by zero
for param in spsa['params']:
a = param['a'] / (spsa['A'] + iter_local) ** spsa['alpha']
c = param['c'] / iter_local ** spsa['gamma']
R = a / c ** 2
flip = 1 if bool(random.getrandbits(1)) else -1
result['w_params'].append({
'name': param['name'],
'value': self.spsa_param_clip_round(param, c * flip, spsa['clipping'], spsa['rounding']),
'R': R,
'c': c,
'flip': flip,
})
result['b_params'].append({
'name': param['name'],
'value': self.spsa_param_clip_round(param, -c * flip, spsa['clipping'], spsa['rounding']),
})
return result
def update_spsa(self, worker, run, spsa_results):
spsa = run['args']['spsa']
if 'clipping' not in spsa:
spsa['clipping'] = 'old'
spsa['iter'] += int(spsa_results['num_games'] / 2)
# Store the history every 'freq' iterations.
# More tuned parameters result in a lower update frequency,
# so that the required storage (performance) remains constant.
if 'param_history' not in spsa:
spsa['param_history'] = []
L = len(spsa['params'])
freq = L * 25
if freq < 100:
freq = 100
maxlen = 250000 / freq
grow_summary = len(spsa['param_history']) < min(maxlen, spsa['iter'] / freq)
# Update the current theta based on the results from the worker
# Worker wins/losses are always in terms of w_params
result = spsa_results['wins'] - spsa_results['losses']
summary = []
w_params = self.get_params(run['_id'], worker)
for idx, param in enumerate(spsa['params']):
R = w_params[idx]['R']
c = w_params[idx]['c']
flip = w_params[idx]['flip']
param['theta'] = self.spsa_param_clip_round(param, R * c * result * flip, spsa['clipping'], 'deterministic')
if grow_summary:
summary.append({
'theta': param['theta'],
'R': R,
'c': c,
})
if grow_summary:
spsa['param_history'].append(summary)
class RunDb:
def __init__(self, db_name='fishtest_new'):
# MongoDB server is assumed to be on the same machine, if not user should use
# ssh with port forwarding to access the remote host.
self.conn = MongoClient(os.getenv('FISHTEST_HOST') or 'localhost')
self.db = self.conn[db_name]
self.userdb = UserDb(self.db)
self.actiondb = ActionDb(self.db)
self.regressiondb = RegressionDb(self.db)
self.runs = self.db['runs']
self.old_runs = self.db['old_runs']
self.chunk_size = 1000
def build_indices(self):
self.runs.ensure_index([('finished', ASCENDING),
('last_updated', DESCENDING)])
def generate_tasks(self, num_games):
tasks = []
remaining = num_games
while remaining > 0:
task_size = min(self.chunk_size, remaining)
tasks.append({
'num_games': task_size,
'pending': True,
'active': False,
})
remaining -= task_size
return tasks
def new_run(self,
base_tag,
new_tag,
num_games,
tc,
book,
book_depth,
threads,
base_options,
new_options,
info='',
resolved_base='',
resolved_new='',
msg_base='',
msg_new='',
base_signature='',
new_signature='',
regression_test=False,
start_time=None,
sprt=None,
spsa=None,
username=None,
tests_repo=None,
auto_purge=True,
throughput=1000,
priority=0):
if start_time == None:
start_time = datetime.utcnow()
run_args = {
'base_tag': base_tag,
'new_tag': new_tag,
'num_games': num_games,
'tc': tc,
'book': book,
'book_depth': book_depth,
'threads': threads,
'regression_test': regression_test,
'resolved_base': resolved_base,
'resolved_new': resolved_new,
'msg_base': msg_base,
'msg_new': msg_new,
'base_options': base_options,
'new_options': new_options,
'info': info,
'base_signature': base_signature,
'new_signature': new_signature,
'username': username,
'tests_repo': tests_repo,
'auto_purge': auto_purge,
'throughput': throughput,
'priority': priority,
'internal_priority': -time.mktime(start_time.timetuple()),
}
if sprt != None:
run_args['sprt'] = sprt
if spsa != None:
run_args['spsa'] = spsa
new_run = {
'args': run_args,
'start_time': start_time,
'last_updated': start_time,
# Will be filled in by tasks, indexed by task-id
'tasks': self.generate_tasks(num_games),
# Aggregated results
'results': {
'wins': 0,
'losses': 0,
'draws': 0
},
'results_stale': False,
'finished': False,
'approved': False,
'approver': '',
}
# Check for an existing approval matching the git commit SHAs
def get_approval(sha):
q = {
'$or': [{
'args.resolved_base': sha
}, {
'args.resolved_new': sha
}],
'approved': True
}
return self.runs.find_one(q)
base_approval = get_approval(resolved_base)
new_approval = get_approval(resolved_new)
allow_auto = username in [
'mcostalba', 'jkiiski', 'glinscott', 'lbraesch'
]
if base_approval != None and new_approval != None and allow_auto:
new_run['approved'] = True
new_run['approver'] = new_approval['approver']
return self.runs.insert(new_run)
def get_machines(self):
machines = []
for run in self.runs.find({'tasks': {'$elemMatch': {'active': True}}}):
for task in run['tasks']:
if task['active']:
machine = copy.copy(task['worker_info'])
machine['last_updated'] = task.get('last_updated', None)
machine['run'] = run
machine['nps'] = task.get('nps', 0)
# TODO(glinscott): Temporary - remove once worker version >= 41
if not isinstance(machine['uname'], basestring):
machine['uname'] = machine['uname'][0] + machine[
'uname'][2]
machines.append(machine)
return machines
# Cache runs
run_cache = {}
run_cache_lock = threading.Lock()
timer = None
def get_run(self, id):
with self.run_cache_lock:
id = str(id)
if id in self.run_cache:
self.run_cache[id]['rtime'] = time.time()
return self.run_cache[id]['run']
run = self.runs.find_one({'_id': ObjectId(id)})
if not run:
run = self.old_runs.find_one({'_id': ObjectId(id)})
self.run_cache[id] = {
'rtime': time.time(),
'ftime': time.time(),
'run': run,
'dirty': False
}
return run
def buffer(self, run, flush):
with self.run_cache_lock:
if self.timer is None:
self.timer = threading.Timer(1.0, self.flush_buffers)
self.timer.start()
id = str(run['_id'])
if flush:
self.run_cache[id] = {
'dirty': False,
'rtime': time.time(),
'ftime': time.time(),
'run': run
}
self.runs.save(run)
else:
if id in self.run_cache:
ftime = self.run_cache[id]['ftime']
else:
ftime = time.time()
self.run_cache[id] = {
'dirty': True,
'rtime': time.time(),
'ftime': ftime,
'run': run
}
def flush_buffers(self):
with self.run_cache_lock:
now = time.time()
old = now + 1
oldest = None
for id in self.run_cache.keys():
if not self.run_cache[id]['dirty']:
if self.run_cache[id]['rtime'] < now - 10:
del self.run_cache[id]
elif self.run_cache[id]['ftime'] < old:
old = self.run_cache[id]['ftime']
oldest = id
if not oldest is None:
if int(now) % 60 == 0:
self.scavenge(self.run_cache[oldest]['run'])
self.runs.save(self.run_cache[oldest]['run'])
self.run_cache[oldest]['dirty'] = False
self.run_cache[oldest]['ftime'] = time.time()
self.timer = threading.Timer(1.0, self.flush_buffers)
self.timer.start()
def scavenge(self, run):
old = datetime.utcnow() - timedelta(minutes=30)
for task in run['tasks']:
if task['active'] and task['last_updated'] < old:
task['active'] = False
def get_run_to_build(self):
return self.runs.find_one({
'binaries_url': {
'$exists': False
},
'finished': False,
'deleted': {
'$exists': False
}
})
def get_runs(self):
return list(self.get_unfinished_runs()) + self.get_finished_runs()[0]
def get_unfinished_runs(self):
with self.run_lock:
return self.runs.find({'finished': False},
sort=[('last_updated', DESCENDING),
('start_time', DESCENDING)])
def get_finished_runs(self,
skip=0,
limit=0,
username='',
success_only=False):
q = {'finished': True, 'deleted': {'$exists': False}}
if len(username) > 0:
q['args.username'] = username
if success_only:
# This is unfortunate, but the only way we have of telling if a run was successful or
# not currently is the color!
q['results_info.style'] = '#44EB44'
c = self.runs.find(q,
skip=skip,
limit=limit,
sort=[('last_updated', DESCENDING)])
result = [list(c), c.count()]
if limit != 0 and len(result[0]) != limit:
c = self.old_runs.find(q,
skip=max(0, skip - c.count()),
limit=limit - len(result[0]))
result[0] += list(c)
result[1] += c.count()
else:
result[1] += self.old_runs.find(q).count()
return result
def get_results(self, run, save_run=True):
if not run['results_stale']:
return run['results']
results = {
'wins': 0,
'losses': 0,
'draws': 0,
'crashes': 0,
'time_losses': 0
}
for task in run['tasks']:
if 'stats' in task:
stats = task['stats']
results['wins'] += stats['wins']
results['losses'] += stats['losses']
results['draws'] += stats['draws']
results['crashes'] += stats['crashes']
results['time_losses'] += stats.get('time_losses', 0)
if 'sprt' in run['args'] and 'state' in run['args']['sprt']:
results['sprt'] = run['args']['sprt']['state']
run['results_stale'] = False
run['results'] = results
if save_run:
self.buffer(run, True)
return results
# Limit concurrent request_task
task_lock = threading.Lock()
task_time = 0
task_runs = None
worker_task = {}
def request_task(self, worker_info):
with self.task_lock:
return self.sync_request_task(worker_info)
def sync_request_task(self, worker_info):
# Check for blocked user or ip
if self.userdb.is_blocked(worker_info):
return {'task_waiting': False}
if time.time() > self.task_time + 60:
self.task_runs = []
for r in self.get_unfinished_runs():
self.task_runs.append(r)
self.task_runs.sort(key=lambda r: (-r['args']['priority'], -r[
'args']['internal_priority'], r['_id']))
self.task_time = time.time()
max_threads = int(worker_info['concurrency'])
exclusion_list = []
# Does this worker have a task already? If so, just hand that back
for runt in self.task_runs:
key = worker_info['unique_key']
if key in self.worker_task and self.worker_task[key] == runt['_id']:
run = self.get_run(runt['_id'])
task_id = -1
for task in run['tasks']:
task_id = task_id + 1
if task['active'] and task['worker_info'] == worker_info:
if task['pending']:
return {'run': run, 'task_id': task_id}
else:
# Don't hand back tasks that have been marked as no longer pending
task['active'] = False
self.buffer(run, True)
# We need to allocate a new task, but first check we don't have the same
# machine already running because multiple connections are not allowed.
connections = 0
for run in self.task_runs:
for task in run['tasks']:
if task['active'] and task['worker_info'][
'remote_addr'] == worker_info['remote_addr']:
connections = connections + 1
# Allow a few connections, for multiple computers on same IP
if connections >= self.userdb.get_machine_limit(
worker_info['username']):
return {'task_waiting': False, 'hit_machine_limit': True}
# Ok, we get a new task that does not require more threads than available concurrency
run_found = False
for runt in self.task_runs:
run = self.get_run(runt['_id'])
if run['_id'] not in exclusion_list and run['approved']:
task_id = -1
for task in run['tasks']:
task_id = task_id + 1
if not task['active'] and task['pending'] and run['args'][
'threads'] <= max_threads:
task['active'] = True
task['last_updated'] = datetime.utcnow()
task['worker_info'] = worker_info
run_found = True
break
if run_found:
break
if not run_found:
return {'task_waiting': False}
# Recalculate internal priority based on task start date and throughput
# Formula: - second_since_epoch - played_and_allocated_tasks * 3600 * chunk_size / games_throughput
# With default value 'throughput = 1000', this means that the priority is unchanged as long as we play at rate '1000 games / hour'.
if (run['args']['throughput'] != None
and run['args']['throughput'] != 0):
run['args']['internal_priority'] = -time.mktime(
run['start_time'].timetuple(
)) - task_id * 3600 * self.chunk_size * run['args'][
'threads'] / run['args']['throughput']
self.buffer(run, True)
self.worker_task[worker_info['unique_key']] = run['_id']
return {'run': run, 'task_id': task_id}
# Create a lock for each active run
run_lock = threading.Lock()
active_runs = {}
purge_count = 0
def active_run_lock(self, id):
with self.run_lock:
self.purge_count = self.purge_count + 1
if self.purge_count > 100000:
old = time.time() - 10000
self.active_runs = dict(
(k, v) for k, v in self.active_runs.iteritems()
if v['time'] >= old)
self.purge_count = 0
if id in self.active_runs:
active_lock = self.active_runs[id]['lock']
self.active_runs[id]['time'] = time.time()
else:
active_lock = threading.Lock()
self.active_runs[id] = {
'time': time.time(),
'lock': active_lock
}
return active_lock
def update_task(self, run_id, task_id, stats, nps, spsa):
lock = self.active_run_lock(str(run_id))
with lock:
return self.sync_update_task(run_id, task_id, stats, nps, spsa)
def sync_update_task(self, run_id, task_id, stats, nps, spsa):
run = self.get_run(run_id)
if task_id >= len(run['tasks']):
return {'task_alive': False}
task = run['tasks'][task_id]
if not task['active'] or not task['pending']:
return {'task_alive': False}
# Guard against incorrect results
count_games = lambda d: d['wins'] + d['losses'] + d['draws']
num_games = count_games(stats)
old_num_games = count_games(
task['stats']) if 'stats' in task else num_games
spsa_games = count_games(spsa) if 'spsa' in run['args'] else 0
if num_games < old_num_games \
or (spsa_games > 0 and num_games <= 0) \
or (spsa_games > 0 and 'stats' in task and num_games <= old_num_games):
return {'task_alive': False}
flush = False
task['stats'] = stats
task['nps'] = nps
if num_games >= task['num_games']:
task['active'] = False
task['pending'] = False
flush = True
update_time = datetime.utcnow()
task['last_updated'] = update_time
run['last_updated'] = update_time
run['results_stale'] = True
# Update spsa results
if 'spsa' in run['args'] and spsa_games == spsa['num_games']:
self.update_spsa(run, spsa)
# Check if SPRT stopping is enabled
if 'sprt' in run['args']:
sprt = run['args']['sprt']
sprt_stats = stat_util.SPRT(self.get_results(run, False),
elo0=sprt['elo0'],
alpha=sprt['alpha'],
elo1=sprt['elo1'],
beta=sprt['beta'],
drawelo=sprt['drawelo'])
if sprt_stats['finished']:
run['args']['sprt']['state'] = sprt_stats['state']
self.stop_run(run_id, run)
flush = True
if (not 'spsa' in run['args'] or spsa_games == spsa['num_games']
or num_games >= task['num_games']
or len(spsa['w_params']) < 20):
self.buffer(run, flush)
return {'task_alive': task['active']}
def failed_task(self, run_id, task_id):
run = self.get_run(run_id)
if task_id >= len(run['tasks']):
return {'task_alive': False}
task = run['tasks'][task_id]
if not task['active'] or not task['pending']:
return {'task_alive': False}
# Mark the task as inactive: it will be rescheduled
task['active'] = False
self.buffer(run, True)
return {}
def stop_run(self, run_id, run=None):
save_it = False
if run is None:
run = self.get_run(run_id)
save_it = True
prune_idx = len(run['tasks'])
for idx, task in enumerate(run['tasks']):
is_active = task['active']
task['active'] = False
task['pending'] = False
if 'stats' not in task and not is_active:
prune_idx = min(idx, prune_idx)
else:
prune_idx = idx + 1
# Truncate the empty tasks
if prune_idx < len(run['tasks']):
del run['tasks'][prune_idx:]
if save_it:
self.buffer(run, True)
self.task_time = 0
return {}
def approve_run(self, run_id, approver):
run = self.get_run(run_id)
# Can't self approve
if run['args']['username'] == approver:
return False
run['approved'] = True
run['approver'] = approver
self.buffer(run, True)
self.task_time = 0
return True
def spsa_param_clip_round(self, param, increment, clipping, rounding):
value = 0.0
if clipping == 'old':
value = min(max(param['theta'] + increment, param['min']),
param['max'])
else: #clipping == 'careful':
inc = min(abs(increment),
abs(param['theta'] - param['min']) / 2,
abs(param['theta'] - param['max']) / 2)
if inc > 0:
inc_sgn = 0 if increment == 0 else increment / abs(increment)
value = param['theta'] + inc_sgn * inc
else: #revert to old behavior to bounce off boundary
value = min(max(param['theta'] + increment, param['min']),
param['max'])
#'deterministic' rounding calls round() inside the worker.
#'randomized' says 4.p should be 5 with probability p, 4 with probability 1-p,
# and is continuous (albeit after expectation) unlike round().
if rounding == 'randomized':
fl = math.floor(
value) #greatest integer <= value, thus works for negative.
if random.uniform(0, 1) < value - fl:
value = fl + 1
else:
value = fl
return value
def request_spsa(self, run_id, task_id):
run = self.get_run(run_id)
if task_id >= len(run['tasks']):
return {'task_alive': False}
task = run['tasks'][task_id]
if not task['active'] or not task['pending']:
return {'task_alive': False}
result = {
'task_alive': True,
'w_params': [],
'b_params': [],
}
spsa = run['args']['spsa']
if 'clipping' not in spsa:
spsa['clipping'] = 'old'
if 'rounding' not in spsa:
spsa['rounding'] = 'deterministic'
# Generate the next set of tuning parameters
iter_local = spsa[
'iter'] + 1 #assume at least one completed, and avoid division by zero
for param in spsa['params']:
a = param['a'] / (spsa['A'] + iter_local)**spsa['alpha']
c = param['c'] / iter_local**spsa['gamma']
R = a / c**2
flip = 1 if bool(random.getrandbits(1)) else -1
result['w_params'].append({
'name':
param['name'],
'value':
self.spsa_param_clip_round(param, c * flip, spsa['clipping'],
spsa['rounding']),
'R':
R,
'c':
c,
'flip':
flip,
})
result['b_params'].append({
'name':
param['name'],
'value':
self.spsa_param_clip_round(param, -c * flip, spsa['clipping'],
spsa['rounding']),
})
return result
def update_spsa(self, run, spsa_results):
spsa = run['args']['spsa']
if 'clipping' not in spsa:
spsa['clipping'] = 'old'
spsa['iter'] += int(spsa_results['num_games'] / 2)
# Update the current theta based on the results from the worker
# Worker wins/losses are always in terms of w_params
result = spsa_results['wins'] - spsa_results['losses']
summary = []
for idx, param in enumerate(spsa['params']):
R = spsa_results['w_params'][idx]['R']
c = spsa_results['w_params'][idx]['c']
flip = spsa_results['w_params'][idx]['flip']
param['theta'] = self.spsa_param_clip_round(
param, R * c * result * flip, spsa['clipping'],
'deterministic')
summary.append({
'theta': param['theta'],
'R': R,
'c': c,
})
# Every 100/1000 iterations, record the parameters
# and stop at 500000 or 200000
if 'param_history' not in spsa:
spsa['param_history'] = []
if len(spsa['params']) < 20:
freq = 100
maxlen = 5001
elif len(spsa['params']) < 50:
freq = 1000
maxlen = 201
else:
freq = 10000
maxlen = 41
if len(spsa['param_history']) < maxlen:
if len(spsa['param_history']) < spsa['iter'] / freq:
spsa['param_history'].append(summary)
elif len(spsa['param_history']) > maxlen:
spsa['param_history'] = spsa['param_history'][-maxlen:]
