def encrypt_all_keys_in_prefix(bucket, prefix, encrypt_out, strip_string,
use_pywren):
keys = utils.list_all_keys(prefix)
print(keys)
if (use_pywren):
chunked_keys = utils.chunks(keys, 500)
def pywren_job(key_chunk):
for key in key_chunk:
utils.encrypt_s3_copy_key(key, bucket, encrypt_out,
strip_string)
return 0
config = wc.default()
config['runtime']['s3_bucket'] = 'imagenet2datav2'
config['runtime'][
's3_key'] = 'pywren.runtime/pywren_runtime-3.6-imagenet2.tar.gz'
pwex = pywren.default_executor(config=config)
print(f"Submitting jobs for {len(keys)} keys")
futures = pwex.map(pywren_job,
chunked_keys,
exclude_modules=["site-packages/"])
pywren.wait(futures)
[f.result() for f in futures]
else:
for key in keys:
utils.encrypt_s3_copy_key(key, bucket, encrypt_out, strip_string)
def reduce(self,
function,
list_of_futures,
extra_env=None,
extra_meta=None):
"""
Apply a function across all futures.
# FIXME change to lazy iterator
"""
#if self.invoker.TIME_LIMIT:
wait(list_of_futures,
return_when=ALL_COMPLETED) # avoid race condition
def reduce_func(fut_list):
# FIXME speed this up for big reduce
accum_list = []
for f in fut_list:
accum_list.append(f.result())
return function(accum_list)
return self.call_async(reduce_func,
list_of_futures,
extra_env=extra_env,
extra_meta=extra_meta)
def map(self, func, iterables):
import pywren
futures = self.pywren_executor.map(
func, iterables, exclude_modules=self.exclude_modules
)
pywren.wait(futures, return_when=pywren.ALL_COMPLETED)
# Set throw_except to False, since when writing Zarr we don't return anything.
# However, Pywren should be improved to distinguish between no return value and an exception.
results = [f.result(throw_except=False) for f in futures]
if self.record_job_history:
run_statuses = [f.run_status for f in futures]
invoke_statuses = [f.invoke_status for f in futures]
outdict = {
"futures": futures,
"run_statuses": run_statuses,
"invoke_statuses": invoke_statuses,
}
logs_dir = os.path.expanduser("~/.zappy/logs")
os.makedirs(logs_dir, exist_ok=True)
timestamp = datetime.datetime.now().strftime("%Y%m%d-%H%M%S.%f")
filename = os.path.join(logs_dir, "pywren-{}.pickle".format(timestamp))
with open(filename, "wb") as file:
pickle.dump(outdict, file)
return results
def test_cholesky_lambda():
X = np.random.randn(128, 128)
A = X.dot(X.T) + np.eye(X.shape[0])
shard_size = 128
shard_sizes = (shard_size, shard_size)
A_sharded = BigMatrix("job_runner_test",
shape=A.shape,
shard_sizes=shard_sizes,
write_header=True)
A_sharded.free()
shard_matrix(A_sharded, A)
program, meta = cholesky(A_sharded)
executor = fs.ProcessPoolExecutor(1)
print("starting program")
program.start()
pwex = pywren.default_executor()
futures = pwex.map(
lambda x: job_runner.lambdapack_run(
program, timeout=60, idle_timeout=6), range(16))
pywren.wait(futures)
print("RESULTSSS")
print([f.result() for f in futures])
futures = pwex.map(
lambda x: job_runner.lambdapack_run(
program, timeout=60, idle_timeout=6), range(16))
program.wait()
#program.free()
L_sharded = meta["outputs"][0]
L_npw = L_sharded.numpy()
L = np.linalg.cholesky(A)
assert (np.allclose(L_npw, L))
print("great success!")
def gemm_recompute(A, B, thresh, s3_key):
"""
Compute A * B.T via speculative execution (i.e., recompute straggling workers).
Params
======
A : numpywren.matrix.BigMatrix
First input matrix.
B : numpywren.matrix.BigMatrix
Second input matrix.
thresh : float (in [0, 1])
Fraction of workers that should finish before recomputing.
s3_key : str
Storage key for output matrix.
Returns
=======
C : matrix.BigMatrix
Resultant matrix product.
t_comp : float
Time for thresh percentage of the workers to finish.
t_straggle : float
Time for the remaining 1 - thresh percentage of the workers to finish after
we begin recomputing.
"""
if not (0 <= thresh <= 1):
raise ValueError("thresh must be in the interval [0, 1]")
"""Initialize output matrix"""
num_col_blocks = A.shape[1] // A.shard_sizes[1]
shard_sizes = (A.shard_sizes[0], B.shard_sizes[0])
C = matrix.BigMatrix(s3_key, shape=(A.shape[0], B.shape[0]), shard_sizes=shard_sizes, autosqueeze=False, write_header=True)
C.delete() # Only needed if you reuse the same s3_key (if the blocks already exist, no work will be done here)
"""Stage 1: Compute "thresh" percentage of the results"""
t_comp_start = time.time()
pwex = pywren.lambda_executor()
futures, num_done = pwex.map(lambda x: pywren_gemm(x, A, B, C, num_col_blocks), C.block_idxs), 0
while num_done < thresh * len(futures):
fs_dones, _ = pywren.wait(futures, return_when=ANY_COMPLETED)
num_done = len(fs_dones)
t_comp = time.time() - t_comp_start # Total stage 1 time
"""Stage 2: Recompute straggling workers (the last 1-thresh percent of jobs)"""
t_straggle_start = time.time()
futures_stragglers = pwex.map(lambda x: pywren_gemm(x, A, B, C, num_col_blocks), C.block_idxs_not_exist)
while len(C.block_idxs_not_exist) > 0:
pywren.wait(futures, return_when=ALWAYS)
pywren.wait(futures_stragglers, return_when=ALWAYS)
t_straggle = time.time() - t_straggle_start # Total stage 2 time
return C, t_comp, t_straggle
def batch_candidates(candidates,
batch_size,
prefix='imagenet2candidates_batches',
bucket='imagenet2datav2',
verbose=True):
import pywren
from pywren import wrenconfig as wc
def hash_ids(ids):
return hashlib.sha256((','.join(ids)).encode()).hexdigest()
def create_and_store_batch(batch_ids):
batch_key = hash_ids(batch_ids)
full_key = os.path.join(prefix, batch_key) + '.pickle'
data = {}
for cur_id in batch_ids:
cur_key = 'imagenet2candidates_scaled/' + cur_id + '.jpg'
data[cur_id], _ = utils.get_s3_object_bytes_with_backoff(cur_key, bucket=bucket)
client = utils.get_s3_client()
client.put_object(Key=full_key, Bucket=bucket, Body=pickle.dumps(data))
return (batch_key, batch_ids)
candidates = candidates.copy()
cids = [c['id_ours'] for c in candidates]
batches = list(utils.chunks(cids, batch_size))
num_batches = len(batches)
if verbose:
print('Creating {} batches of size {} ...'.format(num_batches, batch_size))
pywren_config = wc.default()
pywren_config['runtime']['s3_bucket'] = 'imagenet2pywren'
pywren_config['runtime']['s3_key'] = 'pywren.runtime/pywren_runtime-3.6-imagenet2pywren.tar.gz'
pwex = pywren.default_executor(config=pywren_config)
print(f"Mapping over {len(batches)} images")
futures = pwex.map(create_and_store_batch,
batches,
exclude_modules=["site-packages"])
ALWAYS = 3
done, not_done = pywren.wait(futures, ALWAYS)
while len(not_done) > 0:
done, not_done = pywren.wait(futures, ALWAYS)
time.sleep(1)
print('done')
result = {}
for res in done:
actual_res = res.result()
for cid in actual_res[1]:
result[cid] = actual_res[0]
print(len(result))
for cand in candidates:
assert cand['id_ours'] in result
cand['batch'] = result[cand['id_ours']]
return candidates
def main():
initialize_db()
wren_executor = pywren.default_executor()
start_time = time.time()
with open(ARTICLES_FILE_PATH, encoding=ENCODING) as articles_file:
data = json.loads(articles_file.read())
futures = wren_executor.map(create_nodes, data['titles'])
pywren.wait(futures)
futures = wren_executor.map(create_edges, data['titles'])
pywren.wait(futures)
end_time = time.time()
print('Duration: {}'.format(end_time - start_time))
def argmin(pwex, X, out_bucket=None, tasks_per_job=1):
futures = pwex.map(lambda x: _argmin_remote(x, X), X.block_idxs)
pywren.wait(futures)
results = [f.result() for f in futures]
if (axis == None):
groups = [(None, results)]
else:
groups = itertools.groupby(sorted(results, key=itemgetter(axis)), key=itemgetter(0))
results = []
for _, group in groups:
group = list(group)
argmins = np.concatenate([g[1] for g in group], axis=axis)
argminmin = np.argmin(np.vstack([g[2] for g in group]), axis=axis)
results.append(argmins[argminmin, :])
return np.hstack(results)
def test_map_doublewait(self):
"""
Make sure we can call wait on a list of futures twice
"""
def plus_one(x):
return x + 1
N = 10
x = np.arange(N)
futures = self.wrenexec.map(plus_one, x)
pywren.wait(futures)
pywren.wait(futures)
res = np.array([f.result() for f in futures])
np.testing.assert_array_equal(res, x + 1)
def reshard_down(bigm, breakdowns, pwex=None):
''' Return a new bigm whose shard sizes are bigm.shard_sizes/break_downs
if a pwex is provided reshards in parallel, else reshards locally (very slow)
This will essentially break down a single block in bigm into several evenly sized sub blocks, breakdowns is a list of integers detailing how much to break a given dimension down into. breakdowns = [2,2] would break each dimension down into 2 independent block so a 4 x 4 block would be replaced by 4, 2 x 2 blocks.
'''
for x,y in zip(bigm.shard_sizes, breakdowns):
assert x % y == 0
new_shard_sizes = [int(x/y) for x,y in zip(bigm.shard_sizes, breakdowns)]
X_sharded_new = BigMatrix("reshard({0},{1})".format(bigm.key, breakdowns), bucket=bigm.bucket, shape=bigm.shape, shard_sizes=new_shard_sizes)
chunked_idxs = []
chunked_absolute_idxs = []
for i in range(len(bigm.shape)):
chunked_idxs.append([tuple(x) for x in matrix_utils.chunk(X_sharded_new._block_idxs(i), breakdowns[i])])
chunked_absolute_idxs.append([tuple(x) for x in matrix_utils.chunk(X_sharded_new._blocks(i), breakdowns[i])])
idxs = [bigm._block_idxs(i) for i in range(len(bigm.shape))]
all_idxs_new = list(itertools.product(*chunked_idxs))
all_idxs_old = list(itertools.product(*idxs))
all_idxs_new_absolute = list(itertools.product(*chunked_absolute_idxs))
idx_info = list(zip(all_idxs_new, all_idxs_old, all_idxs_new_absolute))
def reshard_func(bidx_info, bigm, bigm_new):
idxs_new, idx_old, idx_absolute = bidx_info
data = bigm.get_block(*idx_old)
logical = list(itertools.product(*idxs_new))
absolute = list(itertools.product(*idx_absolute))
offsets = [x[0][0] for x in idx_absolute]
for lidx, aidx in zip(logical, absolute):
aidx_offsets = [ slice(x[0] - ox, x[1] - ox) for x,ox in zip(aidx, offsets)]
sub_data = data.__getitem__(aidx_offsets)
print(lidx, aidx, idx_old)
bigm_new.put_block(sub_data, *lidx)
if (pwex is None):
[reshard_func(x, bigm, X_sharded_new) for x in idx_info]
else:
futures = pwex.map(lambda x: reshard_func(x, bigm, X_sharded_new), idx_info)
pywren.wait(futures)
[f.result() for f in futures]
return X_sharded_new
def pytry(fn, arg, DEBUG=False):
# Script to attempt a pywren job
tries = 10
results = []
while tries > 0:
try:
pwex = pywren.default_executor()
futures = pwex.map(fn, arg)
dones, not_dones = pywren.wait(futures, pywren.ALL_COMPLETED)
results = [f.result() for f in dones]
except Exception as e:
raise
if DEBUG:
pickle.dump({
"e": e,
"futures": futures
}, open("debug.pickle", 'wb'))
print('Pickle')
return None
print('oops')
tries -= 1
else:
print('OK')
return results
print('NOT OK')
return results
def benchmark(loopcount, workers, matn, verbose=False):
t1 = time.time()
N = workers
iters = np.arange(N)
def f(x):
return {'flops': compute_flops(loopcount, matn)}
pwex = pywren.lambda_executor(shard_runtime=True)
futures = pwex.map(f, iters)
print("invocation done, dur=", time.time() - t1)
print("callset id: ", futures[0].callset_id)
local_jobs_done_timeline = []
result_count = 0
while result_count < N:
fs_dones, fs_notdones = pywren.wait(futures, pywren.wren.ALWAYS)
result_count = len(fs_dones)
local_jobs_done_timeline.append((time.time(), result_count))
est_flop = 2 * result_count * loopcount * matn**3
est_gflops = est_flop / 1e9 / (time.time() - t1)
if verbose:
print("jobs done: {:5d} runtime: {:5.1f}s {:8.1f} GFLOPS ".
format(result_count,
time.time() - t1, est_gflops))
if result_count == N:
break
time.sleep(1)
if verbose:
print("getting results")
results = [f.result(throw_except=False) for f in futures]
if verbose:
print("getting status")
run_statuses = [f.run_status for f in futures]
invoke_statuses = [f.invoke_status for f in futures]
all_done = time.time()
total_time = all_done - t1
print("total time", total_time)
est_flop = result_count * 2 * loopcount * matn**3
print est_flop / 1e9 / total_time, "GFLOPS"
res = {
'total_time': total_time,
'est_flop': est_flop,
'run_statuses': run_statuses,
'invoke_statuses': invoke_statuses,
'callset_id': futures[0].callset_id,
'local_jobs_done_timeline': local_jobs_done_timeline,
'results': results
}
return res
def benchmark(outfile, workers, experiments, eta, sleep):
experiment_data = []
for exp_i in range(experiments):
print "running experiment {} ---------------------------".format(exp_i)
N = workers
t1 = time.time()
iters = np.arange(N)
def fingerprint(x):
timing_responses = {}
for server in exampleutils.NTP_SERVERS:
ts_os = exampleutils.get_time_offset(server, 4)
timing_responses[server] = ts_os
hwaddr = exampleutils.get_hwaddr()
uptime = exampleutils.get_uptime()[0]
time.sleep(sleep)
return {
'hw_addr': hwaddr,
'ntp_offsets': timing_responses,
'uptime': uptime
}
pwex = pywren.default_executor()
futures = pwex.map(fingerprint, iters)
print "invocation done, dur=", time.time() - t1
print "callset id: ", futures[0].callset_id
fs_dones, fs_notdones = pywren.wait(futures)
# get the job state of all of them
print len(fs_dones), len(fs_notdones)
for f in futures:
if f._state == pywren.wren.JobState.success or f._state == pywren.wren.JobState.error:
pass
else:
print f._state
print "getting results"
results = [f.result() for f in futures]
print "getting status"
run_statuses = [f._run_status for f in futures]
t2 = time.time()
total_time = t2 - t1
sleep_for = eta**exp_i
exp = {
'total_time': total_time,
'exp_i': exp_i,
'sleep_for': sleep_for,
'run_statuses': run_statuses,
'callset_id': futures[0].callset_id,
'results': results
}
experiment_data.append(exp)
print "seeping for", sleep_for / 60.0, "min"
time.sleep(sleep_for)
pickle.dump(experiment_data, open(outfile, 'w'), -1)
def test_numba(self):
"""
Simple numba test, compares two loops, makes sure
one runs much faster than the other
"""
N = 10000000
results = self.wrenexec.map(time_foo, [N])
pywren.wait(results)
regular_time = results[0].result()
print('regular time', regular_time)
results = self.wrenexec.map(time_bar, [N])
pywren.wait(results)
numba_time = results[0].result()
print('numba time', numba_time)
speed_gain = regular_time / numba_time
self.assertTrue(speed_gain > 8.0)
def start(self, parallel=False):
put(self.control_plane.client, self.hash, PS.RUNNING.value)
print("len starters", len(self.program.starters))
chunked_starters = chunk(self.program.starters, 100)
def start_chunk(c):
sqs = boto3.resource('sqs')
queue = sqs.Queue(self.queue_urls[0])
for x in c:
self.set_node_status(*x, NS.READY)
queue.send_message(MessageBody=json.dumps(
[x[0], {str(key): val
for key, val in x[1].items()}]))
if (parallel):
pwex = pywren.default_executor()
futures = pwex.map(start_chunk, chunked_starters)
pywren.wait(futures)
else:
for c in chunked_starters:
start_chunk(c)
return 0
def map(self, func, iterables):
import pywren
futures = self.pywren_executor.map(
func, iterables, exclude_modules=self.exclude_modules)
pywren.wait(futures, return_when=pywren.ALL_COMPLETED)
results = [f.result() for f in futures]
if self.record_job_history:
run_statuses = [f.run_status for f in futures]
invoke_statuses = [f.invoke_status for f in futures]
outdict = {
"futures": futures,
"run_statuses": run_statuses,
"invoke_statuses": invoke_statuses,
}
logs_dir = os.path.expanduser("~/.zappy/logs")
os.makedirs(logs_dir, exist_ok=True)
timestamp = datetime.datetime.now().strftime("%Y%m%d-%H%M%S.%f")
filename = os.path.join(logs_dir,
"pywren-{}.pickle".format(timestamp))
with open(filename, "wb") as file:
pickle.dump(outdict, file)
return results
def test_all_complete(self):
def wait_x_sec_and_plus_one(x):
time.sleep(x)
return x + 1
N = 10
x = np.arange(N)
futures = pywren.default_executor().map(wait_x_sec_and_plus_one, x)
fs_dones, fs_notdones = pywren.wait(futures,
return_when=pywren.wren.ALL_COMPLETED)
res = np.array([f.result() for f in fs_dones])
np.testing.assert_array_equal(res, x+1)
def test_multiple_callset_id_diff_executors(self):
def wait_x_sec_and_plus_one(x):
time.sleep(x)
return x + 1
N = 10
x = np.arange(N)
futures1 = pywren.default_executor().map(wait_x_sec_and_plus_one, x)
futures2 = pywren.default_executor().map(wait_x_sec_and_plus_one, x)
fs_dones, fs_notdones = pywren.wait(futures1 + futures2,
return_when=pywren.wren.ALL_COMPLETED)
res = np.array([f.result() for f in fs_dones])
np.testing.assert_array_equal(res, np.concatenate((x,x))+1)
def wait_for_futures(futures, print_frequency=100, raise_exception=True):
results = []
retrieved = {}
call_id_to_failed_future = {}
done, not_dones = pywren.wait(futures, ALWAYS)
while len(not_dones) > 0:
done, not_dones = pywren.wait(futures, ALWAYS)
for finished_future in done:
#print('finished future')
if finished_future not in retrieved:
try:
if "stdout" in finished_future.run_status:
if len(finished_future.run_status["stdout"]) > 0:
print(finished_future.run_status["stdout"])
#print('Adding finished future to result')
results.append(finished_future.result())
retrieved[finished_future] = True
if len(retrieved) % print_frequency == 0:
timing_results = [
timing_info for _, timing_info in results
]
summarize_timing_infos(timing_results)
except:
if finished_future._traceback is not None:
print("Future exception traceback was ",
finished_future._traceback)
print('future call_id {} failed'.format(
finished_future.call_id))
retrieved[finished_future] = True
call_id_to_failed_future[int(
finished_future.call_id)] = finished_future
if raise_exception:
reraise(finished_future._traceback[0],
finished_future._traceback[1],
finished_future._traceback[2])
return results, call_id_to_failed_future
def test_map(self):
def plus_one(x):
return x + 1
N = 10
x = np.arange(N)
futures = self.wrenexec.map(plus_one, x)
result_count = 0
while result_count < N:
fs_dones, fs_notdones = pywren.wait(futures)
result_count = len(fs_dones)
res = np.array([f.result() for f in futures])
np.testing.assert_array_equal(res, x + 1)
def coded_mat_vec_mul(A_coded_2D, x, num_parity_blocks, coding_length):
def coded_mat_vec_mul(id):
shard_size = A_coded_2D.shard_sizes[1]
reduce_idxs = A_coded_2D._block_idxs(axis=1)
x_loc = x.get_block(0, 0)
Ax_block = None
for r in reduce_idxs:
block1 = A_coded_2D.get_block(id, r)
sidx = r * shard_size
eidx = (r + 1) * shard_size
x_block = x_loc[sidx:eidx]
if (Ax_block is None):
Ax_block = block1.dot(x_block)
else:
Ax_block = Ax_block + block1.dot(x_block)
return Ax_block
shard_size = A_coded_2D.shard_sizes[0]
n_coded_procs = len(A_coded_2D._block_idxs(0))
len_A_coded = n_coded_procs - coding_length - 1
pwex = pywren.lambda_executor()
futures = pwex.map(coded_mat_vec_mul, range(n_coded_procs))
Ax = np.zeros((A_coded_2D.shape[0], 1))
bitmask = np.ones((coding_length + 1, num_parity_blocks + 1))
not_done = list(range(n_coded_procs))
while decode2D.cant_be_decoded_with_bitmask(deepcopy(bitmask)):
fs_dones, fs_not_dones = pywren.wait(futures, 2)
for (id, f) in enumerate(futures):
if f in fs_dones and id in not_done:
# print("Worker done", id)
try:
Ax[id * shard_size:(id + 1) * shard_size] = f.result()
i, j = decode2D.ind1Dto2D(id, len_A_coded,
num_parity_blocks)
bitmask[i, j] = 0
not_done.remove(id)
except Exception as e:
print(e)
pass
# print("1: Decoding not dones", not_done)
Ax = decode2D.decode_vector_with_bitmask(Ax, bitmask, num_parity_blocks,
len_A_coded, shard_size,
coding_length)
return Ax
def test_any_complete(self):
def wait_x_sec_and_plus_one(x):
time.sleep(x)
return x + 1
N = 10
x = np.arange(N)
futures = pywren.default_executor().map(wait_x_sec_and_plus_one, x)
fs_notdones = futures
while (len(fs_notdones) > 0):
fs_dones, fs_notdones = pywren.wait(fs_notdones,
return_when=pywren.wren.ANY_COMPLETED,
WAIT_DUR_SEC=1)
self.assertTrue(len(fs_dones) > 0)
res = np.array([f.result() for f in futures])
np.testing.assert_array_equal(res, x+1)
def progwait(fs, desc="", notebook=False):
"""
A tqdm-based progress bar that looks nice and gives you an
idea of how long until your futures are done
"""
N = len(fs)
result_count = 0
fs_dones = []
fs_notdones = fs
if notebook:
from tqdm import tqdm_notebook as tqdm_func
else:
from tqdm import tqdm as tqdm_func
with tqdm_func(total=N, desc=desc) as pbar:
while len(fs_dones) < N:
new_fs_dones, new_fs_notdones = pywren.wait(
fs_notdones, return_when=pywren.ANY_COMPLETED)
fs_dones += new_fs_dones
fs_notdones = new_fs_notdones
pbar.update(len(new_fs_dones))
return fs_dones, fs_notdones
def process_pending(self):
active_map_ids = list(self.active_iters.keys())
logger.info("ppc={} begin process pending len_active_map_ids={}".format(self._process_pending_count, len(active_map_ids)))
for map_id in active_map_ids:
logger.debug("processing map_id={} {}".format(map_id, len(self.active_iters[map_id])))
# get a future set
iter_futures = self.active_iters[map_id]
# group these by callset ID
f_by_callset_id = {}
for f in iter_futures:
pwf = f.current_future
if pwf.callset_id not in f_by_callset_id:
f_by_callset_id[pwf.callset_id] = []
f_by_callset_id[pwf.callset_id].append(pwf)
## call WAIT on everyone
logger.debug("map_id={} starting status check".format(map_id))
for cs, flist in f_by_callset_id.items():
logger.debug("map_id={} calling wait for callset_id={} len_futures={}".format(map_id,cs, len(flist)))
# this will trigger an update on all of them
fs_done, fs_notdone = pywren.wait(flist, return_when=pywren.ALWAYS, # ANY_COMPLETED,
WAIT_DUR_SEC=1)
logger.debug("map_id={} wait done for callset_id={} len_fs_done={}".format(map_id,cs, len(fs_done)))
logger.debug("map_id={} status check done for all f in map_id".format(map_id))
to_advance = []
to_remove = []
still_waiting = []
for f in iter_futures:
pwf = f.current_future
if f.current_iter == f.max_iter:
to_remove.append(f)
else:
if pwf.succeeded():
to_advance.append(f)
elif pwf.errored():
logger.debug("map_id={} map_pos={} errored on iter {}".format(map_id,
f.original_map_pos,
f.current_iter))
to_remove.append(f)
else:
still_waiting.append(f)
logger.debug("map_id={} to_advance={}".format(map_id, get_map_pos(to_advance)))
logger.debug("map_id={} to_remove={}".format(map_id, get_map_pos(to_remove)))
logger.debug("map_id={} still_waiting={}".format(map_id, get_map_pos(still_waiting)))
if len(to_advance) > 0:
# construct next invocation
wrapped_args = [(f.current_iter + 1,
f.current_future.result(),
f.arg) for f in to_advance ]
wrapped_func = self.wrapped_funcs[map_id]
logger.debug("map_id={} invoking new map with {} args".format(map_id, len(wrapped_args)))
pywren_futures = self.wrenexec.map(wrapped_func,
wrapped_args,
exclude_modules=EXCLUDE_MODULES)
logger.debug("map_id={} invoking new map done".format(map_id))
for f, pwf in zip(to_advance, pywren_futures):
if f.save_iters:
f.iter_hist.append(f.current_future)
f.current_future = pwf
f.current_iter += 1
# remove these from current map id
to_remove_map_pos = [f.original_map_pos for f in to_remove]
self.active_iters[map_id] = [f for f in self.active_iters[map_id] if f.original_map_pos \
not in to_remove_map_pos]
if len(self.active_iters[map_id]) == 0:
logger.debug("map_id={} deleted".format(map_id))
del self.active_iters[map_id]
del self.wrapped_funcs[map_id]
logger.info("ppc={} end process pending".format(self._process_pending_count))
self._process_pending_count += 1
def code_2D(A, num_parity_blocks, thres=1):
assert (len(A._block_idxs(0)) % num_parity_blocks == 0)
shard_size = A.shard_sizes[0]
coded_shape = (A.shape[0] + num_parity_blocks * A.shard_sizes[0],
A.shape[1])
coding_length = int(np.ceil(len(A._block_idxs(0)) / num_parity_blocks))
coding_fn2D = make_coding_function2D(A, coding_length)
coded_2D_shape = (
A.shape[0] +
(coding_length + 1 + num_parity_blocks) * A.shard_sizes[0], A.shape[1])
A_coded_2D = matrix.BigMatrix(A.key + "CODED2D_{0}_{1}_{2}".format(
A.shape[0], shard_size, num_parity_blocks),
shape=coded_2D_shape,
shard_sizes=A.shard_sizes,
write_header=True,
parent_fn=coding_fn2D)
# if list(set(A_coded_2D.block_idxs_not_exist) - set(A.block_idxs_exist)) == []:
# return A_coded_2D
last_block = max(A._block_idxs(0))
columns = A_coded_2D._block_idxs(1)
rows = A_coded_2D._block_idxs(0)
to_read = []
blocks_exist = A_coded_2D.block_idxs_exist
for row in rows:
if (row <= last_block): continue
for column in columns:
if (row, column) in blocks_exist:
continue
else:
to_read.append((row, column))
print("Number of parity blocks", len(to_read))
num_parities_1D = coding_length * len(A._block_idxs(1))
to_read_phase1 = to_read[0:num_parities_1D]
to_read_phase2 = to_read[num_parities_1D:]
def get_block_wrapper(x):
A_coded_2D.get_block(*x)
return 0
#### For 2D ENCODING of A, uncomment
pwex = pywren.lambda_executor()
t_enc1 = time.time()
futures2 = pwex.map(get_block_wrapper, to_read_phase1)
result_count = 0
fs_dones = []
while (result_count < thres * len(to_read_phase1)):
fs_dones, fs_notdones = pywren.wait(futures2, 2)
result_count = len(fs_dones)
print(result_count)
time.sleep(3)
for f in fs_dones:
try:
f.result()
except Exception as e:
print(e)
pass
t_enc1 = time.time() - t_enc1
print("Encoding phase 1 time", t_enc1)
t_enc2 = time.time()
futures2 = pwex.map(get_block_wrapper, to_read_phase2)
result_count = 0
while (result_count < thres * len(to_read_phase2)):
fs_dones, fs_notdones = pywren.wait(futures2, 2)
result_count = len(fs_dones)
print(result_count)
time.sleep(3)
for f in fs_dones:
try:
f.result()
except Exception as e:
print(e)
pass
t_enc2 = time.time() - t_enc2
print("Encoding phase 2 time", t_enc2)
print("Total ENCODING time", t_enc1 + t_enc2)
# a = list(set(A_coded_2D.block_idxs_not_exist) - set(A.block_idxs_exist))
# print("Still to encode", a)
return A_coded_2D
def write_data():
def run_command(key):
"""
keylist.append({'taskId': i,
'job_number': job_number,
'total_input': numTasks,
'write_element_size': write_element_size,
'process_time': process_time,
'total_time': total_time,
'redis': redisnode})
"""
pywren.wrenlogging.default_config('INFO')
begin_of_function = time.time()
logger = logging.getLogger(__name__)
logger.info("taskId = " + str(key['taskId']))
taskId = key['taskId']
jobid_int = int(key['job_number'])
write_element_size = int(key['write_element_size'])
process_time = int(key['process_time'])
total_time = int(key['total_time'])
rs = []
#for hostname in key['redis'].split(";"):
# r1 = StrictRedis(host=hostname, port=6379, db=0).pipeline()
# rs.append(r1)
#r1 = StrictRedis(host="172.31.12.131", port=6379, db=0).pipeline()
#rs.append(r1)
#nrs = len(rs)
nrs = 1
[read_time, work_time, write_time] = [0] * 3
start_time = time.time()
# a total of 10 threads
number_of_clients = 1
write_pool = ThreadPool(number_of_clients)
time.sleep(process_time)
logger.info("Process finish here: " + str(time.time()))
def write_work_client(writer_key):
start_time = time.time()
client_id = int(writer_key['client_id'])
taskID = writer_key['taskId']
jobID = writer_key['jobid']
datasize = writer_key['write_element_size']
#datasize = 1310720
total_time = writer_key['total_time']
body = b'a' * datasize
client_id = int(client_id)
count = 0
throughput_step = 1
throughput_count = 1
throughput_total = 0
throughput_nops = 0
ret = []
while time.time() < start_time + total_time:
count = count + 1
keyname = str(jobID) + "-" + str(taskID) + "-" + str(count)
m = hashlib.md5()
m.update(keyname.encode('utf-8'))
ridx = int(m.hexdigest()[:8], 16) % nrs
randomized_keyname = str(jobID) + "-" + str(
taskID) + '-' + m.hexdigest()[:8] + '-' + str(count)
#logger.info("(" + str(taskId) + ")" + "The name of the key to write is: " + randomized_keyname)
start = time.time()
logger.info("[REDIS] [" + str(jobID) + "] " +
str(time.time()) + " " + str(taskID) + " " +
str(len(body)) + " write " + "S")
#rs[ridx].set(randomized_keyname, body)
end = time.time()
logger.info("[REDIS] [" + str(jobID) + "] " +
str(time.time()) + " " + str(taskID) + " " +
str(len(body)) + " write " + "E ")
#for r in rs:
# r.execute()
throughput_total += end - start
throughput_nops += 1
if end - start_time >= throughput_count:
throughput = throughput_nops / throughput_total
ret.append((end, throughput))
throughput_nops = 0
throughput_count += throughput_step
throughput_total = 0
logger.info("Write finish here: " + str(time.time()))
return ret
writer_keylist = []
number_of_clients = int(number_of_clients)
for i in range(number_of_clients):
writer_keylist.append({
'client_id': i,
'taskId': taskId,
'jobid': jobid_int,
'write_element_size': write_element_size,
'total_time': total_time
})
start_time = time.time()
write_pool_handler_container = []
write_pool_handler = write_pool.map_async(write_work_client,
writer_keylist)
write_pool_handler_container.append(write_pool_handler)
if len(write_pool_handler_container) > 0:
write_pool_handler = write_pool_handler_container.pop()
ret = write_pool_handler.get()
twait_end = time.time()
write_time = twait_end - start_time
write_pool.close()
write_pool.join()
end_of_function = time.time()
return begin_of_function, end_of_function, write_time, ret
numTasks = int(sys.argv[1])
job_number = int(sys.argv[2])
write_element_size = int(sys.argv[3])
process_time = int(sys.argv[4]) # microseconds
total_time = int(sys.argv[5])
redisnode = sys.argv[6]
keylist = []
for i in range(numTasks):
keylist.append({
'taskId': i,
'job_number': job_number,
'total_input': numTasks,
'write_element_size': write_element_size,
'process_time': process_time,
'total_time': total_time,
'redis': redisnode
})
wrenexec = pywren.default_executor()
futures = wrenexec.map(run_command, keylist)
pywren.wait(futures)
results = [f.result() for f in futures]
print("Write " + str(job_number))
run_statuses = [f.run_status for f in futures]
invoke_statuses = [f.invoke_status for f in futures]
res = {
'results': results,
'run_statuses': run_statuses,
'invoke_statuses': invoke_statuses
}
filename = "redis-write-" + ".pickle.breakdown"
pickle.dump(res, open(filename, 'wb'))
return res
MAT_N = 4096
iters = np.arange(N)
def f(x):
return compute_flops(LOOPCOUNT, MAT_N)
pwex = pywren.default_executor()
futures = pwex.map(f, iters)
print "invocation done, dur=", time.time() - t1
print futures[0].callset_id
result_count = 0
while result_count < N:
fs_dones, fs_notdones = pywren.wait(futures)
result_count = len(fs_dones)
est_flop = 2 * result_count * LOOPCOUNT * MAT_N**3
est_gflops = est_flop / 1e9 / (time.time() - t1)
print "jobs done: {:5d} runtime: {:5.1f}s {:8.1f} GFLOPS ".format(
result_count,
time.time() - t1, est_gflops)
if result_count == N:
break
time.sleep(1)
all_done = time.time()
"/tmp/gensort", "-b" + str(begin),
str(number_of_records), "/dev/stdout"
])
keyname = "input/part-" + str(key)
m = hashlib.md5()
m.update(keyname.encode('utf-8'))
logging.info("Here is the key! " + str(key))
randomized_keyname = "input/" + m.hexdigest()[:8] + "-part-" + str(
key)
put_start = time.time()
client.put_object(Body=data,
Bucket="yupeng-pywren-0",
Key=randomized_keyname)
put_end = time.time()
logger.info(
str(key) + " th object uploaded using " +
str(put_end - put_start) + " seconds.")
gc.collect()
key = key + 1
wrenexec = pywren.default_executor()
num_of_files = int(sys.argv[1])
print("Generating " + str(62.5 * int(num_of_files)) + " Mb input dataset.")
passed_tasks = range(0, num_of_files, 5)
fut = wrenexec.map(run_command, passed_tasks)
pywren.wait(fut)
res = [f.result() for f in fut]
print(res)
def gemm(pwex,
X,
Y,
out_bucket=None,
tasks_per_job=1,
local=False,
dtype=np.float64,
overwrite=True,
gemm_impl=0,
gemm_chunk_size=16):
'''
Compute XY return
@param pwex - Execution context
@param X - rhs matrix
@param Y - lhs matrix
@param tasks_per_job - number of tasks per job
@param out_bucket - bucket job writes to
@param num_jobs - how many lambdas to run
@param local - run locally? #TODO remove once local pywren executor is provided
'''
# 0 -> 1 or 1 -> 0
reduce_idxs = Y._block_idxs(axis=0)
if (out_bucket == None):
out_bucket = X.bucket
root_key = generate_key_name_binop(X, Y, "gemm")
if (Y.shard_sizes[0] != X.shard_sizes[1]):
raise Exception("X dim 1 shard size must match Y dim 0 shard size")
XY = BigMatrix(root_key,
shape=(X.shape[0], Y.shape[1]),
bucket=out_bucket,
shard_sizes=[X.shard_sizes[0], Y.shard_sizes[1]],
dtype=dtype,
write_header=True)
num_out_blocks = len(XY.blocks)
if (tasks_per_job > num_out_blocks):
tasks_per_job = 1
num_jobs = int(num_out_blocks / float(tasks_per_job))
print("Out Shape", XY.shape)
print("Total number of output blocks", len(XY.block_idxs))
print("Total number of output blocks that exist", len(XY.blocks_exist))
if (overwrite):
block_idxs_to_map = list(set(XY.block_idxs))
else:
block_idxs_to_map = list(set(XY.block_idxs_not_exist))
print("Number of output blocks to generate ", len(block_idxs_to_map))
chunked_blocks = list(
chunk(list(chunk(block_idxs_to_map, tasks_per_job)), num_jobs))
if (not isinstance(pwex.invoker, pywren.queues.SQSInvoker)
and gemm_impl > 0):
raise Exception(
"GEMM IMPL > 0 only supported for standalone mode pywren")
print(_gemms[gemm_impl])
def pywren_run(x):
return _gemms[gemm_impl](x,
XY,
X,
Y,
reduce_idxs=reduce_idxs,
dtype=dtype,
block_chunk_size=gemm_chunk_size)
all_futures = []
for i, c in enumerate(chunked_blocks):
print("Submitting job for chunk {0} in axis 0".format(i))
if (local):
list(map(pywren_run, c))
else:
s = time.time()
futures = pwex.map(pywren_run,
c,
exclude_modules=["site-packages"])
e = time.time()
print("Pwex Map Time {0}".format(e - s))
all_futures.append((i, futures))
if (local):
return XY
for i, futures, in all_futures:
print("waiting")
pywren.wait(futures)
[f.result() for f in futures]
return XY
def recompute_mat_vec_mul(A_coded_2D, x, thres=0.95):
def shard_mat_vec_mul(id):
shard_size = A_coded_2D.shard_sizes[1]
reduce_idxs = A_coded_2D._block_idxs(axis=1)
x_loc = x.get_block(0, 0)
Ax_block = None
for r in reduce_idxs:
block1 = A_coded_2D.get_block(id, r)
sidx = r * shard_size
eidx = (r + 1) * shard_size
x_block = x_loc[sidx:eidx]
if (Ax_block is None):
Ax_block = block1.dot(x_block)
else:
Ax_block = Ax_block + block1.dot(x_block)
return Ax_block, id
shard_size = A_coded_2D.shard_sizes[0]
n_coded_procs = len(A_coded_2D._block_idxs(0))
pwex = pywren.lambda_executor()
futures = pwex.map(shard_mat_vec_mul, range(n_coded_procs))
Ax = np.zeros((A_coded_2D.shape[0], 1))
not_done = list(range(n_coded_procs))
fs_dones = []
f_result_dones = []
while len(fs_dones) < thres * n_coded_procs:
fs_dones, fs_not_dones = pywren.wait(futures, 2)
for f in list(set(fs_dones) - set(f_result_dones)):
# print("Worker done", id)
f_result_dones.append(f)
try:
result = f.result()
id = result[1]
Ax[id * shard_size:(id + 1) * shard_size] = result[0]
not_done.remove(id)
except Exception as e:
#print(e)
pass
time.sleep(2)
print("Recomputing not dones", not_done)
futures2 = pwex.map(shard_mat_vec_mul, not_done)
f_result_dones2 = []
while not_done != []:
fs_dones2, fs_not_dones2 = pywren.wait(futures2, 3)
for f in list(set(fs_dones2) - set(f_result_dones2)):
f_result_dones2.append(f)
try:
result = f.result()
id = result[1]
if id in not_done:
print("Recomputed", id)
Ax[id * shard_size:(id + 1) * shard_size] = result[0]
not_done.remove(id)
except Exception as e:
#print(e)
pass
time.sleep(2)
fs_dones, fs_not_dones = pywren.wait(futures, 3)
for f in list(set(fs_dones) - set(f_result_dones)):
f_result_dones.append(f)
try:
result = f.result()
id = result[1]
if id in not_done:
print("Straggler computed", id)
Ax[id * shard_size:(id + 1) * shard_size] = result[0]
not_done.remove(id)
except Exception as e:
#print(e)
pass
time.sleep(2)
if fs_not_dones2 == [] and fs_not_dones == []:
print("NOT DONE", not_done)
break
print("Recomputing done")
return Ax
