def test_netflix_sgd(ctx):
U = 100
M = 100 * 100
r = 20
d = 8
P_RATING = 1000.0 / (U * M)
# create random factor and value matrices
Mfactor = spartan.eager(spartan.rand(M, r).astype(np.float32))
Ufactor = spartan.eager(spartan.rand(U, r).astype(np.float32))
V = spartan.sparse_empty((U, M),
tile_hint=(divup(U, d), divup(M, d)),
dtype=np.float32)
# V = spartan.shuffle(V, netflix.load_netflix_mapper,
# kw={ 'load_file' : '/big1/netflix.zip' })
V = spartan.eager(
spartan.tocoo(
spartan.shuffle(V,
netflix.fake_netflix_mapper,
target=V,
kw={'p_rating': P_RATING})))
for i in range(2):
_ = netflix.sgd(V, Mfactor, Ufactor).force()
def test_netflix_sgd(ctx):
U = 100
M = 100*100
r = 20
d = 8
P_RATING = 1000.0 / (U * M)
# create random factor and value matrices
Mfactor = spartan.eager(spartan.rand(M, r).astype(np.float32))
Ufactor = spartan.eager(spartan.rand(U, r).astype(np.float32))
V = spartan.sparse_empty((U, M),
tile_hint=(divup(U, d), divup(M, d)),
dtype=np.float32)
# V = spartan.shuffle(V, netflix.load_netflix_mapper,
# kw={ 'load_file' : '/big1/netflix.zip' })
V = spartan.eager(
spartan.tocoo(
spartan.shuffle(V, netflix.fake_netflix_mapper,
target=V, kw={'p_rating': P_RATING})))
for i in range(2):
_ = netflix.sgd(V, Mfactor, Ufactor).evaluate()
def benchmark_convnet(ctx, timer):
image_size = BASE_IMG_SIZE
minibatch = 64
#minibatch = ctx.num_workers
hint = util.divup(image_size, sqrt(ctx.num_workers))
tile_hint = (util.divup(minibatch, ctx.num_workers), N_COLORS, image_size, image_size)
util.log_info('Hint: %s', tile_hint)
images = expr.eager(expr.ones((minibatch, N_COLORS, image_size, image_size),
tile_hint=tile_hint))
w1 = expr.eager(expr.ones((N_FILTERS, N_COLORS) + FILTER_SIZE,
tile_hint=ONE_TILE))
w2 = expr.eager(expr.ones((N_FILTERS, N_FILTERS) + FILTER_SIZE,
tile_hint=ONE_TILE))
w3 = expr.eager(expr.ones((N_FILTERS, N_FILTERS) + FILTER_SIZE,
tile_hint=ONE_TILE))
def _():
conv1 = stencil.stencil(images, w1, 2)
pool1 = stencil.maxpool(conv1)
conv2 = stencil.stencil(pool1, w2, 2)
pool2 = stencil.maxpool(conv2)
conv3 = stencil.stencil(pool2, w3, 2)
pool3 = stencil.maxpool(conv3)
expr.force(pool3)
# force parakeet functions to compile before timing.
_()
for i in range(2):
timer.time_op('convnet', _)
def benchmark_netflix_sgd(ctx, timer):
d = ctx.num_workers
V = spartan.sparse_empty((U, M),
tile_hint=(divup(U, d), divup(M, d)),
dtype=np.float32)
V = timer.time_op(
'prep', lambda: spartan.eager(
spartan.tocoo(
spartan.shuffle(V,
netflix.fake_netflix_mapper,
target=V,
kw={'p_rating': P_RATING}))))
# V = spartan.shuffle(V, netflix.load_netflix_mapper,
# kw={ 'load_file' : '/big1/netflix.zip' })
for r in [25, 50]:
Mfactor = spartan.eager(
spartan.rand(M, r, tile_hint=(divup(M, d), r)).astype(np.float32))
Ufactor = spartan.eager(
spartan.rand(U, r, tile_hint=(divup(U, d), r)).astype(np.float32))
timer.time_op('rank %d' % r, netflix.sgd(V, Mfactor, Ufactor).force)
def benchmark_convnet(ctx, timer):
image_size = BASE_IMG_SIZE
minibatch = 64
#minibatch = ctx.num_workers
hint = util.divup(image_size, sqrt(ctx.num_workers))
tile_hint = (util.divup(minibatch, ctx.num_workers), N_COLORS, image_size, image_size)
util.log_info('Hint: %s', tile_hint)
images = expr.eager(expr.ones((minibatch, N_COLORS, image_size, image_size),
tile_hint=tile_hint))
w1 = expr.eager(expr.ones((N_FILTERS, N_COLORS) + FILTER_SIZE,
tile_hint=ONE_TILE))
w2 = expr.eager(expr.ones((N_FILTERS, N_FILTERS) + FILTER_SIZE,
tile_hint=ONE_TILE))
w3 = expr.eager(expr.ones((N_FILTERS, N_FILTERS) + FILTER_SIZE,
tile_hint=ONE_TILE))
def _():
conv1 = stencil.stencil(images, w1, 2)
pool1 = stencil.maxpool(conv1)
conv2 = stencil.stencil(pool1, w2, 2)
pool2 = stencil.maxpool(conv2)
conv3 = stencil.stencil(pool2, w3, 2)
pool3 = stencil.maxpool(conv3)
pool3.evaluate()
# force parakeet functions to compile before timing.
_()
for i in range(2):
timer.time_op('convnet', _)
def test_convnet(ctx):
hint = util.divup(64, sqrt(ctx.num_workers))
images = expr.eager(
expr.ones((N_IMGS, ) + IMG_SIZE,
tile_hint=(N_IMGS, N_COLORS, hint, hint)))
w1 = expr.eager(
expr.ones((N_FILTERS, N_COLORS) + FILTER_SIZE, tile_hint=ONE_TILE))
conv1 = stencil.stencil(images, w1, 2)
pool1 = stencil.maxpool(conv1)
w2 = expr.eager(
expr.ones((N_FILTERS, N_FILTERS) + FILTER_SIZE, tile_hint=ONE_TILE))
conv2 = stencil.stencil(pool1, w2, 2)
pool2 = stencil.maxpool(conv2)
w3 = expr.eager(
expr.ones((N_FILTERS, N_FILTERS) + FILTER_SIZE, tile_hint=ONE_TILE))
conv3 = stencil.stencil(pool2, w3, 2)
pool3 = stencil.maxpool(conv3)
util.log_info(pool3.shape)
def test_kmeans_expr(self):
ctx = spartan.blob_ctx.get()
pts = expr.rand(N_PTS, N_DIM,
tile_hint=(divup(N_PTS, ctx.num_workers), N_DIM)).force()
k = KMeans(N_CENTERS, ITER)
k.fit(pts)
def benchmark_svm(ctx, timer):
print "#worker:", ctx.num_workers
max_iter = 2
#N = 200000 * ctx.num_workers
N = 1000 * 64
D = 64
# create data
data = expr.randn(N, D, dtype=np.float64, tile_hint=(N, util.divup(D, ctx.num_workers)))
labels = expr.shuffle(data, _init_label_mapper, shape_hint=(data.shape[0], 1))
t1 = datetime.now()
w = fit(data, labels, T=max_iter).force()
t2 = datetime.now()
util.log_warn('train time per iteration:%s ms, final w:%s', millis(t1,t2)/max_iter, w.glom().T)
correct = 0
for i in range(10):
new_data = expr.randn(1, D, dtype=np.float64, tile_hint=[1, D])
new_label = predict(w, new_data)
#print 'point %s, predict %s' % (new_data.glom(), new_label)
new_data = new_data.glom()
if new_data[0,0] >= new_data[0,1] and new_label == 1.0 or new_data[0,0] < new_data[0,1] and new_label == -1.0:
correct += 1
print 'predict precision:', correct * 1.0 / 10
def precompute(self):
'''Precompute the most k similar items for each item.
After this funcion returns. 2 attributes will be created.
Attributes
------
top_k_similar_table : Numpy array of shape (N, k).
Records the most k similar scores between each items.
top_k_similar_indices : Numpy array of shape (N, k).
Records the indices of most k similar items for each item.
'''
M = self.rating_table.shape[0]
N = self.rating_table.shape[1]
self.similarity_table = expr.shuffle(self.rating_table, _similarity_mapper,
kw={'item_norm': self._get_norm_of_each_item(self.rating_table),
'step': util.divup(self.rating_table.shape[1], blob_ctx.get().num_workers)},
shape_hint=(N, N))
# Release the memory for item_norm
top_k_similar_indices = expr.zeros((N, self.k), dtype=np.int)
# Find top-k similar items for each item.
# Store the similarity scores into table top_k_similar table.
# Store the indices of top k items into table top_k_similar_indices.
cost = np.prod(top_k_similar_indices.shape)
top_k_similar_table = expr.shuffle(self.similarity_table, _select_most_k_similar_mapper,
kw = {'top_k_similar_indices': top_k_similar_indices, 'k': self.k},
shape_hint=(N, self.k),
cost_hint={hash(top_k_similar_indices):{'00': 0, '01': cost, '10': cost, '11': cost}})
self.top_k_similar_table = top_k_similar_table.optimized().glom()
self.top_k_similar_indices = top_k_similar_indices.optimized().glom()
def benchmark_netflix_sgd(ctx, timer):
d = ctx.num_workers
V = spartan.sparse_empty((U, M), tile_hint=(divup(U, d), divup(M, d)), dtype=np.float32)
V = timer.time_op(
"prep",
lambda: spartan.eager(
spartan.tocoo(spartan.shuffle(V, netflix.fake_netflix_mapper, target=V, kw={"p_rating": P_RATING}))
),
)
# V = spartan.shuffle(V, netflix.load_netflix_mapper,
# kw={ 'load_file' : '/big1/netflix.zip' })
for r in [25, 50]:
Mfactor = spartan.eager(spartan.rand(M, r, tile_hint=(divup(M, d), r)).astype(np.float32))
Ufactor = spartan.eager(spartan.rand(U, r, tile_hint=(divup(U, d), r)).astype(np.float32))
timer.time_op("rank %d" % r, netflix.sgd(V, Mfactor, Ufactor).force)
def benchmark_matmul(ctx, timer):
N = int(1000 * math.pow(ctx.num_workers, 1.0 / 3.0))
# N = 4000
M = util.divup(N, ctx.num_workers)
T = util.divup(N, math.sqrt(ctx.num_workers))
util.log_info("Testing with %d workers, N = %d, tile_size=%s", ctx.num_workers, N, T)
# x = expr.eager(expr.ones((N, N), dtype=np.double, tile_hint=(N, M)))
# y = expr.eager(expr.ones((N, N), dtype=np.double, tile_hint=(N, M)))
x = expr.eager(expr.ones((N, N), dtype=np.double, tile_hint=(T, T)))
y = expr.eager(expr.ones((N, N), dtype=np.double, tile_hint=(T, T)))
# print expr.glom(expr.dot(x, y))
# print expr.dag(expr.dot(x, y))
def _step():
expr.evaluate(expr.dot(x, y))
timer.time_op("matmul", _step)
def precompute(self):
'''Precompute the most k similar items for each item.
After this funcion returns. 2 attributes will be created.
Attributes
------
top_k_similar_table : Numpy array of shape (N, k).
Records the most k similar scores between each items.
top_k_similar_indices : Numpy array of shape (N, k).
Records the indices of most k similar items for each item.
'''
M = self.rating_table.shape[0]
N = self.rating_table.shape[1]
self.similarity_table = expr.shuffle(
self.rating_table,
_similarity_mapper,
kw={
'item_norm':
self._get_norm_of_each_item(self.rating_table),
'step':
util.divup(self.rating_table.shape[1],
blob_ctx.get().num_workers)
},
shape_hint=(N, N))
# Release the memory for item_norm
top_k_similar_indices = expr.zeros((N, self.k), dtype=np.int)
# Find top-k similar items for each item.
# Store the similarity scores into table top_k_similar table.
# Store the indices of top k items into table top_k_similar_indices.
cost = np.prod(top_k_similar_indices.shape)
top_k_similar_table = expr.shuffle(self.similarity_table,
_select_most_k_similar_mapper,
kw={
'top_k_similar_indices':
top_k_similar_indices,
'k': self.k
},
shape_hint=(N, self.k),
cost_hint={
hash(top_k_similar_indices): {
'00': 0,
'01': cost,
'10': cost,
'11': cost
}
})
self.top_k_similar_table = top_k_similar_table.optimized().glom()
self.top_k_similar_indices = top_k_similar_indices.optimized().glom()
def benchmark_matmul(ctx, timer):
N = int(1000 * math.pow(ctx.num_workers, 1.0 / 3.0))
#N = 4000
M = util.divup(N, ctx.num_workers)
T = util.divup(N, math.sqrt(ctx.num_workers))
util.log_info('Testing with %d workers, N = %d, tile_size=%s',
ctx.num_workers, N, T)
#x = expr.eager(expr.ones((N, N), dtype=np.double, tile_hint=(N, M)))
#y = expr.eager(expr.ones((N, N), dtype=np.double, tile_hint=(N, M)))
x = expr.eager(expr.ones((N, N), dtype=np.double, tile_hint=(T, T)))
y = expr.eager(expr.ones((N, N), dtype=np.double, tile_hint=(T, T)))
#print expr.glom(expr.dot(x, y))
#print expr.dag(expr.dot(x, y))
def _step():
expr.evaluate(expr.dot(x, y))
timer.time_op('matmul', _step)
def start_cluster(num_workers, use_cluster_workers):
'''
Start a cluster with ``num_workers`` workers.
If use_cluster_workers is True, then use the remote workers
defined in `spartan.config`. Otherwise, workers are all
spawned on the localhost.
:param num_workers:
:param use_cluster_workers:
'''
rpc.set_default_timeout(FLAGS.default_rpc_timeout)
#clean the checkpoint directory
if os.path.exists(FLAGS.checkpoint_path):
shutil.rmtree(FLAGS.checkpoint_path)
master = spartan.master.Master(FLAGS.port_base, num_workers)
ssh_processes = []
if not use_cluster_workers:
start_remote_worker('localhost', 0, num_workers)
else:
available_workers = sum([cnt for _, cnt in FLAGS.hosts])
assert available_workers >= num_workers, 'Insufficient slots to run all workers.'
count = 0
num_hosts = len(FLAGS.hosts)
for worker, total_tasks in FLAGS.hosts:
if FLAGS.assign_mode == AssignMode.BY_CORE:
sz = total_tasks
else:
sz = util.divup(num_workers, num_hosts)
sz = min(sz, num_workers - count)
ssh_processes.append(start_remote_worker(worker, count,
count + sz))
count += sz
if count == num_workers:
break
master.wait_for_initialization()
# Kill the now unnecessary ssh processes.
# Fegin : if we kill these processes, we can't get log from workers.
#for process in ssh_processes:
#process.kill()
return master
def start_cluster(num_workers, use_cluster_workers):
'''
Start a cluster with ``num_workers`` workers.
If use_cluster_workers is True, then use the remote workers
defined in `spartan.config`. Otherwise, workers are all
spawned on the localhost.
:param num_workers:
:param use_cluster_workers:
'''
rpc.set_default_timeout(FLAGS.default_rpc_timeout)
#clean the checkpoint directory
if os.path.exists(FLAGS.checkpoint_path):
shutil.rmtree(FLAGS.checkpoint_path)
master = spartan.master.Master(FLAGS.port_base, num_workers)
ssh_processes = []
if not use_cluster_workers:
start_remote_worker('localhost', 0, num_workers)
else:
available_workers = sum([cnt for _, cnt in FLAGS.hosts])
assert available_workers >= num_workers, 'Insufficient slots to run all workers.'
count = 0
num_hosts = len(FLAGS.hosts)
for worker, total_tasks in FLAGS.hosts:
if FLAGS.assign_mode == AssignMode.BY_CORE:
sz = total_tasks
else:
sz = util.divup(num_workers, num_hosts)
sz = min(sz, num_workers - count)
ssh_processes.append(start_remote_worker(worker, count, count + sz))
count += sz
if count == num_workers:
break
master.wait_for_initialization()
# Kill the now unnecessary ssh processes.
# Fegin : if we kill these processes, we can't get log from workers.
#for process in ssh_processes:
#process.kill()
return master
def benchmark_als(ctx, timer):
print "#worker:", ctx.num_workers
#USER_SIZE = 100 * ctx.num_workers
USER_SIZE = 320
#USER_SIZE = 200 * 64
MOVIE_SIZE = 12800
num_features = 20
num_iter = 2
A = expr.randint(USER_SIZE, MOVIE_SIZE, low=0, high=5, tile_hint=(USER_SIZE, util.divup(MOVIE_SIZE, ctx.num_workers)))
#A = expr.randint(USER_SIZE, MOVIE_SIZE, low=0, high=5)
util.log_warn('begin als!')
t1 = datetime.now()
U, M = als(A, implicit_feedback=True, num_features=num_features, num_iter=num_iter)
U.force()
M.force()
t2 = datetime.now()
cost_time = millis(t1,t2)
print "total cost time:%s ms, per iter cost time:%s ms" % (cost_time, cost_time/num_iter)
def benchmark_pr(ctx, timer):
num_pages = 300 * 1000 * 3 * ctx.num_workers
num_outlinks = 10
density = num_outlinks * 1.0 / num_pages
same_site_prob = 0.9
print "#worker:", ctx.num_workers
col_step = util.divup(num_pages, ctx.num_workers)
wts_tile_hint = [num_pages, col_step]
p_tile_hint = [col_step, 1]
#wts = expr.sparse_diagonal((num_pages, num_pages), dtype=np.float32, tile_hint=wts_tile_hint)
#wts = expr.eager(
# expr.sparse_rand((num_pages, num_pages),
# density=density,
# format='csr',
# dtype=np.float32,
# tile_hint=wts_tile_hint))
wts = pagerank_sparse(num_pages, num_outlinks, same_site_prob)
#res = wts.glom().todense()
#for i in range(res.shape[0]):
# l = []
# for j in range(res.shape[1]):
# l.append(round(res[i,j],1))
# print l
#p = expr.sparse_empty((num_pages,1), dtype=np.float32, tile_hint=p_tile_hint).evaluate()
#for i in range(num_pages):
# p[i,0] = 1
#p = expr.sparse_rand((num_pages, 1), density=1.0, format='csc', dtype=np.float32, tile_hint=p_tile_hint)
p = expr.rand(num_pages, 1).astype(np.float32)
#q = expr.zeros((num_pages, 1), dtype=np.float32, tile_hint=p_tile_hint).evaluate()
#q[:] = p.glom().todense()
#q = expr.lazify(q)
#r = expr.dot(wts, p)
#print r.glom()
t1 = datetime.now()
sparse_multiply(wts, p, p_tile_hint)
t2 = datetime.now()
cost_time = millis(t1, t2)
print 'current benchmark:', cost_time / num_iter / 1000
def benchmark_pr(ctx, timer): num_pages = 300 * 1000 * 3 * ctx.num_workers num_outlinks = 10 density = num_outlinks * 1.0 / num_pages same_site_prob = 0.9 print "#worker:", ctx.num_workers col_step = util.divup(num_pages, ctx.num_workers) wts_tile_hint = [num_pages, col_step] p_tile_hint = [col_step, 1] #wts = expr.sparse_diagonal((num_pages, num_pages), dtype=np.float32, tile_hint=wts_tile_hint) #wts = expr.eager( # expr.sparse_rand((num_pages, num_pages), # density=density, # format='csr', # dtype=np.float32, # tile_hint=wts_tile_hint)) wts = pagerank_sparse(num_pages, num_outlinks, same_site_prob) #res = wts.glom().todense() #for i in range(res.shape[0]): # l = [] # for j in range(res.shape[1]): # l.append(round(res[i,j],1)) # print l #p = expr.sparse_empty((num_pages,1), dtype=np.float32, tile_hint=p_tile_hint).evaluate() #for i in range(num_pages): # p[i,0] = 1 #p = expr.sparse_rand((num_pages, 1), density=1.0, format='csc', dtype=np.float32, tile_hint=p_tile_hint) p = expr.rand(num_pages, 1).astype(np.float32) #q = expr.zeros((num_pages, 1), dtype=np.float32, tile_hint=p_tile_hint).evaluate() #q[:] = p.glom().todense() #q = expr.lazify(q) #r = expr.dot(wts, p) #print r.glom() t1 = datetime.now() sparse_multiply(wts, p, p_tile_hint) t2 = datetime.now() cost_time = millis(t1, t2) print 'current benchmark:', cost_time / num_iter / 1000
def test_stencil(ctx):
st = time.time()
IMG_SIZE = int(8 * math.sqrt(ctx.num_workers))
FILT_SIZE = 8
N = 8
F = 32
tile_size = util.divup(IMG_SIZE, math.sqrt(ctx.num_workers))
images = expr.ones((N, 3, IMG_SIZE, IMG_SIZE),
dtype=np.float,
tile_hint=(N, 3, tile_size, tile_size))
filters = expr.ones((F, 3, FILT_SIZE, FILT_SIZE),
dtype=np.float,
tile_hint=ONE_TILE)
result = stencil.stencil(images, filters, 1)
ed = time.time()
print ed - st
def test_convnet(ctx):
hint = util.divup(64, sqrt(ctx.num_workers))
images = expr.eager(expr.ones((N_IMGS,) + IMG_SIZE,
tile_hint=(N_IMGS, N_COLORS, hint, hint)))
w1 = expr.eager(expr.ones((N_FILTERS, N_COLORS) + FILTER_SIZE,
tile_hint=ONE_TILE))
conv1 = stencil.stencil(images, w1, 2)
pool1 = stencil.maxpool(conv1)
w2 = expr.eager(expr.ones((N_FILTERS, N_FILTERS) + FILTER_SIZE,
tile_hint=ONE_TILE))
conv2 = stencil.stencil(pool1, w2, 2)
pool2 = stencil.maxpool(conv2)
w3 = expr.eager(expr.ones((N_FILTERS, N_FILTERS) + FILTER_SIZE,
tile_hint=ONE_TILE))
conv3 = stencil.stencil(pool2, w3, 2)
pool3 = stencil.maxpool(conv3)
util.log_info(pool3.shape)