def testInitialAssignsWithInputs(self):
import numpy as np
from mars.tensor.random import TensorRandint
from mars.tensor.arithmetic import TensorTreeAdd
n1 = TensorRandint(state=np.random.RandomState(0),
dtype=np.float32()).new_chunk(None, shape=(10, 10))
n2 = TensorRandint(state=np.random.RandomState(1),
dtype=np.float32()).new_chunk(None, shape=(10, 10))
n3 = TensorTreeAdd(dtype=np.float32()).new_chunk(None, shape=(10, 10))
n3.op._inputs = [n1, n2]
n4 = TensorTreeAdd(dtype=np.float32()).new_chunk(None, shape=(10, 10))
n4.op._inputs = [n3]
graph = DAG()
graph.add_node(n1)
graph.add_node(n3)
graph.add_node(n4)
graph.add_edge(n1, n3)
graph.add_edge(n3, n4)
analyzer = GraphAnalyzer(graph, {})
ext_chunks = analyzer.collect_external_input_chunks(initial=False)
self.assertListEqual(ext_chunks[n3.op.key], [n2.key])
self.assertEqual(
len(analyzer.collect_external_input_chunks(initial=True)), 0)
def _build_chunk_dag(node_str, edge_str):
from mars.tensor.random import TensorRandint
from mars.tensor.arithmetic import TensorTreeAdd
char_dag = DAG()
for s in node_str.split(','):
char_dag.add_node(s.strip())
for s in edge_str.split(','):
l, r = s.split('->')
char_dag.add_edge(l.strip(), r.strip())
chunk_dag = DAG()
str_to_chunk = dict()
for s in char_dag.topological_iter():
if char_dag.count_predecessors(s):
c = TensorTreeAdd(args=[],
_key=s,
dtype=np.dtype(np.float32())).new_chunk(
None, shape=(10, 10)).data
inputs = c.op._inputs = [
str_to_chunk[ps] for ps in char_dag.predecessors(s)
]
else:
c = TensorRandint(_key=s, dtype=np.dtype(
np.float32())).new_chunk(None, shape=(10, 10)).data
inputs = []
str_to_chunk[s] = c
chunk_dag.add_node(c)
for inp in inputs:
chunk_dag.add_edge(inp, c)
return chunk_dag, str_to_chunk
async def test_assign_gpu_tasks(actor_pool):
pool, session_id, assigner_ref, cluster_api, meta_api = actor_pool
input1 = TensorFetch(key='a', source_key='a',
dtype=np.dtype(int)).new_chunk([])
input2 = TensorFetch(key='b', source_key='b',
dtype=np.dtype(int)).new_chunk([])
result_chunk = TensorTreeAdd(args=[input1, input2], gpu=True) \
.new_chunk([input1, input2])
chunk_graph = ChunkGraph([result_chunk])
chunk_graph.add_node(input1)
chunk_graph.add_node(input2)
chunk_graph.add_node(result_chunk)
chunk_graph.add_edge(input1, result_chunk)
chunk_graph.add_edge(input2, result_chunk)
await meta_api.set_chunk_meta(input1,
memory_size=200,
store_size=200,
bands=[('address0', 'numa-0')])
await meta_api.set_chunk_meta(input2,
memory_size=200,
store_size=200,
bands=[('address0', 'numa-0')])
subtask = Subtask('test_task', session_id, chunk_graph=chunk_graph)
[result] = await assigner_ref.assign_subtasks([subtask])
assert result[1].startswith('gpu')
def testSameKeyAssign(self):
import numpy as np
from mars.tensor.random import TensorRandint
from mars.tensor.arithmetic import TensorTreeAdd
graph = DAG()
r"""
Proper initial allocation should divide the graph like
U U | U U | U U
| | | | | | | | | | | | | |
U U | U U | U U
"""
inputs = [
tuple(
TensorRandint(_key=str(i), dtype=np.float32()).new_chunk(
None, shape=(10, 10)) for _ in range(2)) for i in range(6)
]
results = [
TensorTreeAdd(dtype=np.float32()).new_chunk(None, shape=(10, 10))
for _ in range(6)
]
for inp, r in zip(inputs, results):
r.op._inputs = list(inp)
graph.add_node(r)
for n in inp:
graph.add_node(n)
graph.add_edge(n, r)
analyzer = GraphAnalyzer(graph, dict(w1=24, w2=24, w3=24))
assignments = analyzer.calc_operand_assignments(
analyzer.get_initial_operand_keys())
self.assertEqual(len(assignments), 6)
async def test_execute_tensor(actor_pool):
pool, session_id, meta_api, storage_api, execution_ref = actor_pool
data1 = np.random.rand(10, 10)
data2 = np.random.rand(10, 10)
input1 = TensorFetch(key='input1',
source_key='input2',
dtype=np.dtype(int)).new_chunk([])
input2 = TensorFetch(key='input2',
source_key='input2',
dtype=np.dtype(int)).new_chunk([])
result_chunk = TensorTreeAdd(args=[input1, input2]) \
.new_chunk([input1, input2], shape=data1.shape, dtype=data1.dtype)
await meta_api.set_chunk_meta(input1,
memory_size=data1.nbytes,
store_size=data1.nbytes,
bands=[(pool.external_address, 'numa-0')])
await meta_api.set_chunk_meta(input2,
memory_size=data1.nbytes,
store_size=data2.nbytes,
bands=[(pool.external_address, 'numa-0')])
# todo use different storage level when storage ready
await storage_api.put(input1.key, data1)
await storage_api.put(input2.key, data2)
chunk_graph = ChunkGraph([result_chunk])
chunk_graph.add_node(input1)
chunk_graph.add_node(input2)
chunk_graph.add_node(result_chunk)
chunk_graph.add_edge(input1, result_chunk)
chunk_graph.add_edge(input2, result_chunk)
subtask = Subtask('test_task',
session_id=session_id,
chunk_graph=chunk_graph)
await execution_ref.run_subtask(subtask, 'numa-0', pool.external_address)
# check if results are correct
result = await storage_api.get(result_chunk.key)
np.testing.assert_array_equal(data1 + data2, result)
# check if quota computations are correct
quota_ref = await mo.actor_ref(QuotaActor.gen_uid('numa-0'),
address=pool.external_address)
[quota] = await quota_ref.get_batch_quota_reqs()
assert quota[(subtask.subtask_id, subtask.subtask_id)] == data1.nbytes
# check if metas are correct
result_meta = await meta_api.get_chunk_meta(result_chunk.key)
assert result_meta['object_id'] == result_chunk.key
assert result_meta['shape'] == result.shape
def testMockExecuteSize(self):
import mars.tensor as mt
from mars.core.graph import DAG
from mars.tensor.fetch import TensorFetch
from mars.tensor.arithmetic import TensorTreeAdd
graph_add = DAG()
input_chunks = []
for _ in range(2):
fetch_op = TensorFetch(dtype=np.dtype('int64'))
inp_chunk = fetch_op.new_chunk(None, shape=(100, 100)).data
input_chunks.append(inp_chunk)
add_op = TensorTreeAdd(args=input_chunks, dtype=np.dtype('int64'))
add_chunk = add_op.new_chunk(input_chunks,
shape=(100, 100),
dtype=np.dtype('int64')).data
graph_add.add_node(add_chunk)
for inp_chunk in input_chunks:
graph_add.add_node(inp_chunk)
graph_add.add_edge(inp_chunk, add_chunk)
executor = Executor()
res = executor.execute_graph(graph_add, [add_chunk.key],
compose=False,
mock=True)[0]
self.assertEqual(res, (80000, 80000))
self.assertEqual(executor.mock_max_memory, 80000)
for _ in range(3):
new_add_op = TensorTreeAdd(args=[add_chunk],
dtype=np.dtype('int64'))
new_add_chunk = new_add_op.new_chunk([add_chunk],
shape=(100, 100),
dtype=np.dtype('int64')).data
graph_add.add_node(new_add_chunk)
graph_add.add_edge(add_chunk, new_add_chunk)
add_chunk = new_add_chunk
executor = Executor()
res = executor.execute_graph(graph_add, [add_chunk.key],
compose=False,
mock=True)[0]
self.assertEqual(res, (80000, 80000))
self.assertEqual(executor.mock_max_memory, 160000)
a = mt.random.rand(10, 10, chunk_size=10)
b = a[:, mt.newaxis, :] - a
r = mt.triu(mt.sqrt(b**2).sum(axis=2))
executor = Executor()
res = executor.execute_tensor(r, concat=False, mock=True)
# larger than maximal memory size in calc procedure
self.assertGreaterEqual(res[0][0], 800)
self.assertGreaterEqual(executor.mock_max_memory, 8000)
def _build_test_graph(data_list):
from mars.tensor.fetch import TensorFetch
from mars.tensor.arithmetic import TensorTreeAdd
inputs = []
for idx, d in enumerate(data_list):
chunk_key = 'chunk-%d' % idx
fetch_chunk = TensorFetch(to_fetch_key=chunk_key, dtype=d.dtype) \
.new_chunk([], shape=d.shape, _key=chunk_key)
inputs.append(fetch_chunk)
add_chunk = TensorTreeAdd(data_list[0].dtype).new_chunk(inputs, shape=data_list[0].shape)
exec_graph = DAG()
exec_graph.add_node(add_chunk)
for input_chunk in inputs:
exec_graph.add_node(input_chunk)
exec_graph.add_edge(input_chunk, add_chunk)
return exec_graph, inputs, add_chunk
def _build_test_graph(data_list):
from mars.tensor.fetch import TensorFetch
from mars.tensor.arithmetic import TensorTreeAdd
inputs = []
for idx, d in enumerate(data_list):
chunk_key = f'chunk-{random.randint(0, 999)}-{idx}'
fetch_chunk = TensorFetch(to_fetch_key=chunk_key, dtype=d.dtype) \
.new_chunk([], shape=d.shape, _key=chunk_key)
inputs.append(fetch_chunk)
add_chunk = TensorTreeAdd(args=inputs, dtype=data_list[0].dtype) \
.new_chunk(inputs, shape=data_list[0].shape)
exec_graph = ChunkGraph([add_chunk.data])
exec_graph.add_node(add_chunk.data)
for input_chunk in inputs:
exec_graph.add_node(input_chunk.data)
exec_graph.add_edge(input_chunk.data, add_chunk.data)
return exec_graph, inputs, add_chunk
async def test_assigner(actor_pool):
pool, session_id, assigner_ref, meta_api = actor_pool
input1 = TensorFetch(key='a', source_key='a',
dtype=np.dtype(int)).new_chunk([])
input2 = TensorFetch(key='b', source_key='b',
dtype=np.dtype(int)).new_chunk([])
input3 = TensorFetch(key='c', source_key='c',
dtype=np.dtype(int)).new_chunk([])
result_chunk = TensorTreeAdd(args=[input1, input2, input3]) \
.new_chunk([input1, input2, input3])
chunk_graph = ChunkGraph([result_chunk])
chunk_graph.add_node(input1)
chunk_graph.add_node(input2)
chunk_graph.add_node(input3)
chunk_graph.add_node(result_chunk)
chunk_graph.add_edge(input1, result_chunk)
chunk_graph.add_edge(input2, result_chunk)
chunk_graph.add_edge(input3, result_chunk)
await meta_api.set_chunk_meta(input1,
memory_size=200,
store_size=200,
bands=[('address0', 'numa-0')])
await meta_api.set_chunk_meta(input2,
memory_size=400,
store_size=400,
bands=[('address1', 'numa-0')])
await meta_api.set_chunk_meta(input3,
memory_size=400,
store_size=400,
bands=[('address2', 'numa-0')])
subtask = Subtask('test_task', session_id, chunk_graph=chunk_graph)
[result] = await assigner_ref.assign_subtasks([subtask])
assert result in (('address1', 'numa-0'), ('address2', 'numa-0'))
def testCompose(self):
"""
test compose in build graph and optimize
"""
r"""
graph(@: node, #: composed_node):
@ --> @ --> @ ========> #
"""
chunks = [
TensorTreeAdd(_key=str(n)).new_chunk(None, None) for n in range(3)
]
graph = DirectedGraph()
lmap(graph.add_node, chunks[:3])
graph.add_edge(chunks[0], chunks[1])
graph.add_edge(chunks[1], chunks[2])
composed_nodes = graph.compose()
self.assertTrue(composed_nodes[0].composed == chunks[:3])
# make the middle one as result chunk, thus the graph cannot be composed
composed_nodes = graph.compose(keys=[chunks[1].key])
self.assertEqual(len(composed_nodes), 0)
r"""
graph(@: node, #: composed_node):
@ @ @ @
\ / \ /
@ --> @ ========> #
/ \ / \
@ @ @ @
"""
chunks = [
TensorTreeAdd(_key=str(n)).new_chunk(None, None) for n in range(6)
]
graph = DirectedGraph()
lmap(graph.add_node, chunks[:6])
chunks[2].op._inputs = [chunks[0], chunks[1]]
chunks[3].op._inputs = [chunks[2]]
chunks[4].op._inputs = [chunks[3]]
chunks[5].op._inputs = [chunks[3]]
graph.add_edge(chunks[0], chunks[2])
graph.add_edge(chunks[1], chunks[2])
graph.add_edge(chunks[2], chunks[3])
graph.add_edge(chunks[3], chunks[4])
graph.add_edge(chunks[3], chunks[5])
composed_nodes = graph.compose()
self.assertTrue(composed_nodes[0].composed == chunks[2:4])
# to make sure the predecessors and successors of compose are right
# 0 and 1's successors must be composed
self.assertIn(composed_nodes[0], graph.successors(chunks[0]))
self.assertIn(composed_nodes[0], graph.successors(chunks[1]))
# check composed's inputs
self.assertIn(chunks[0].key, [n.key for n in composed_nodes[0].inputs])
self.assertIn(chunks[1].key, [n.key for n in composed_nodes[0].inputs])
# check composed's predecessors
self.assertIn(chunks[0], graph.predecessors(composed_nodes[0]))
self.assertIn(chunks[1], graph.predecessors(composed_nodes[0]))
# check 4 and 5's inputs
self.assertIn(
composed_nodes[0].key,
[n.key for n in graph.successors(composed_nodes[0])[0].inputs])
self.assertIn(
composed_nodes[0].key,
[n.key for n in graph.successors(composed_nodes[0])[0].inputs])
# check 4 and 5's predecessors
self.assertIn(composed_nodes[0], graph.predecessors(chunks[4]))
self.assertIn(composed_nodes[0], graph.predecessors(chunks[5]))
# test optimizer compose
r"""
graph(@: node, S: Slice Chunk, #: composed_node):
@ @ @ @
\ / \ /
@ --> @ --> S ========> # --> S
/ \ / \
@ @ @ @
compose stopped at S, because numexpr don't support Slice op
"""
chunks = [
TensorTreeAdd(_key=str(n)).new_chunk(None, None) for n in range(6)
]
chunk_slice = TensorSlice().new_chunk([None], None)
graph = DirectedGraph()
lmap(graph.add_node, chunks[:6])
graph.add_node(chunk_slice)
graph.add_edge(chunks[0], chunks[2])
graph.add_edge(chunks[1], chunks[2])
graph.add_edge(chunks[2], chunks[3])
graph.add_edge(chunks[3], chunk_slice)
graph.add_edge(chunk_slice, chunks[4])
graph.add_edge(chunk_slice, chunks[5])
optimizer = NeOptimizer(graph)
composed_nodes = optimizer.compose()
self.assertTrue(composed_nodes[0].composed == chunks[2:4])
r"""
graph(@: node, S: Slice Chunk, #: composed_node):
@ --> @ --> S --> @ ========> # --> S --> @
compose stopped at S, because numexpr don't support Slice op
"""
chunks = [
TensorTreeAdd(_key=str(n)).new_chunk(None, None) for n in range(4)
]
graph = DirectedGraph()
lmap(graph.add_node, chunks[:3])
graph.add_node(chunk_slice)
graph.add_edge(chunks[0], chunks[1])
graph.add_edge(chunks[1], chunk_slice)
graph.add_edge(chunk_slice, chunks[2])
optimizer = NeOptimizer(graph)
composed_nodes = optimizer.compose()
self.assertTrue(composed_nodes[0].composed == chunks[:2])
self.assertTrue(len(composed_nodes) == 1)
r"""
graph(@: node, S: Slice Chunk, #: composed_node):
@ --> @ --> S --> @ --> @ ========> # --> S --> #
compose stopped at S, because numexpr don't support Slice op
"""
chunks = [
TensorTreeAdd(_key=str(n)).new_chunk(None, None) for n in range(4)
]
graph = DirectedGraph()
lmap(graph.add_node, chunks[:4])
graph.add_node(chunk_slice)
graph.add_edge(chunks[0], chunks[1])
graph.add_edge(chunks[1], chunk_slice)
graph.add_edge(chunk_slice, chunks[2])
graph.add_edge(chunks[2], chunks[3])
optimizer = NeOptimizer(graph)
composed_nodes = optimizer.compose()
self.assertTrue(composed_nodes[0].composed == chunks[:2])
self.assertTrue(composed_nodes[1].composed == chunks[2:4])
def testAssignOnWorkerLost(self):
import numpy as np
from mars.scheduler import OperandState
from mars.tensor.random import TensorRandint
from mars.tensor.arithmetic import TensorTreeAdd
graph = DAG()
r"""
Proper initial allocation should divide the graph like
FL FL F F R R | FL FL F F R R
| x | | x | | x | | | x | | x | | x |
R R R R U U | R R R R U U
U: UNSCHEDULED F: FINISHED R: READY L: LOST
"""
op_states = dict()
inputs = [
tuple(
TensorRandint(
dtype=np.float32()).new_chunk(None, shape=(10, 10))
for _ in range(2)) for _ in range(6)
]
results = [
tuple(
TensorTreeAdd(_key=f'{i}_{j}', dtype=np.float32()).new_chunk(
None, shape=(10, 10)) for j in range(2)) for i in range(6)
]
for inp, outp in zip(inputs, results):
for o in outp:
o.op._inputs = list(inp)
op_states[o.op.key] = OperandState.UNSCHEDULED
graph.add_node(o)
for n in inp:
op_states[n.op.key] = OperandState.UNSCHEDULED
graph.add_node(n)
for o in outp:
graph.add_edge(n, o)
fixed_assigns = dict()
for idx in range(4):
for i in range(2):
fixed_assigns[inputs[idx][i].op.key] = f'w{idx % 2 + 1}'
op_states[inputs[idx][i].op.key] = OperandState.FINISHED
fixed_assigns[results[idx][i].op.key] = f'w{idx % 2 + 1}'
op_states[results[idx][i].op.key] = OperandState.READY
for inp in inputs:
for n in inp:
if n.op.key in fixed_assigns:
continue
op_states[n.op.key] = OperandState.READY
lost_chunks = [c.key for inp in (inputs[0], inputs[2]) for c in inp]
worker_metrics = dict(w2=24, w3=24)
analyzer = GraphAnalyzer(graph, worker_metrics, fixed_assigns,
op_states, lost_chunks)
changed_states = analyzer.analyze_state_changes()
self.assertEqual(len(changed_states), 8)
self.assertTrue(
all(changed_states[c.op.key] == OperandState.READY
for inp in (inputs[0], inputs[2]) for c in inp))
self.assertTrue(
all(changed_states[c.op.key] == OperandState.UNSCHEDULED
for res in (results[0], results[2]) for c in res))
assignments = analyzer.calc_operand_assignments(
analyzer.get_initial_operand_keys())
for inp in inputs:
if any(n.op.key in fixed_assigns for n in inp):
continue
self.assertEqual(1, len(set(assignments[n.op.key] for n in inp)))
worker_assigns = dict((k, 0) for k in worker_metrics)
for w in assignments.values():
worker_assigns[w] += 1
self.assertEqual(2, worker_assigns['w2'])
self.assertEqual(6, worker_assigns['w3'])
def testAssignWithPreviousData(self):
import numpy as np
from mars.scheduler.chunkmeta import WorkerMeta
from mars.tensor.random import TensorRandint
from mars.tensor.arithmetic import TensorTreeAdd
graph = DAG()
r"""
Proper initial allocation should divide the graph like
U U | U U | U U
\ / | \ / | \ /
U | U | U
"""
inputs = [
tuple(
TensorRandint(_key=str(i * 2 +
j), dtype=np.float32()).new_chunk(
None, shape=(10, 10))
for j in range(2)) for i in range(3)
]
results = [
TensorTreeAdd(dtype=np.float32()).new_chunk(None, shape=(10, 10))
for _ in range(3)
]
for inp, r in zip(inputs, results):
r.op._inputs = list(inp)
graph.add_node(r)
for n in inp:
graph.add_node(n)
graph.add_edge(n, r)
# assign with partial mismatch
data_dist = {
'0':
dict(c00=WorkerMeta(chunk_size=5, workers=('w1', )),
c01=WorkerMeta(chunk_size=5, workers=('w2', ))),
'1':
dict(c10=WorkerMeta(chunk_size=10, workers=('w1', ))),
'2':
dict(c20=WorkerMeta(chunk_size=10, workers=('w3', ))),
'3':
dict(c30=WorkerMeta(chunk_size=10, workers=('w3', ))),
'4':
dict(c40=WorkerMeta(chunk_size=7, workers=('w3', ))),
}
analyzer = GraphAnalyzer(graph, dict(w1=24, w2=24, w3=24))
assignments = analyzer.calc_operand_assignments(
analyzer.get_initial_operand_keys(), input_chunk_metas=data_dist)
self.assertEqual(len(assignments), 6)
# explanation of the result:
# for '1', all data are in w1, hence assigned to w1
# '0' assigned to w1 according to connectivity
# '2' and '3' assigned to w3 according to connectivity
# '4' assigned to w2 because it has fewer data, and the slots of w3 is used up
self.assertEqual(assignments['0'], 'w1')
self.assertEqual(assignments['1'], 'w1')
self.assertEqual(assignments['2'], 'w3')
self.assertEqual(assignments['3'], 'w3')
self.assertEqual(assignments['4'], 'w2')
self.assertEqual(assignments['5'], 'w2')
# assign with full mismatch
data_dist = {
'0':
dict(c00=WorkerMeta(chunk_size=5, workers=('w1', )),
c01=WorkerMeta(chunk_size=5, workers=(
'w1',
'w2',
))),
'1':
dict(c10=WorkerMeta(chunk_size=10, workers=('w1', ))),
'2':
dict(c20=WorkerMeta(chunk_size=10, workers=('w3', ))),
'3':
dict(c30=WorkerMeta(chunk_size=10, workers=('w3', ))),
'4':
dict(c40=WorkerMeta(chunk_size=7, workers=('w2', ))),
'5':
dict(c50=WorkerMeta(chunk_size=7, workers=('w2', ))),
}
analyzer = GraphAnalyzer(graph, dict(w1=24, w2=24, w3=24))
assignments = analyzer.calc_operand_assignments(
analyzer.get_initial_operand_keys(), input_chunk_metas=data_dist)
self.assertEqual(len(assignments), 6)
self.assertEqual(assignments['0'], 'w1')
self.assertEqual(assignments['1'], 'w1')
self.assertEqual(assignments['2'], 'w3')
self.assertEqual(assignments['3'], 'w3')
self.assertEqual(assignments['4'], 'w2')
self.assertEqual(assignments['5'], 'w2')
def testAssignOnWorkerAdd(self):
import numpy as np
from mars.scheduler import OperandState
from mars.tensor.random import TensorRandint
from mars.tensor.arithmetic import TensorTreeAdd
graph = DAG()
r"""
Proper initial allocation should divide the graph like
F F R R | F F R R | R R R R
| x | | x | | | x | | x | | | x | | x |
R R U U | R R U U | U U U U
U: UNSCHEDULED F: FINISHED R: READY
"""
inputs = [
tuple(
TensorRandint(
dtype=np.float32()).new_chunk(None, shape=(10, 10))
for _ in range(2)) for _ in range(6)
]
results = [
tuple(
TensorTreeAdd(_key='%d_%d' %
(i, j), dtype=np.float32()).new_chunk(
None, shape=(10, 10)) for j in range(2))
for i in range(6)
]
for inp, outp in zip(inputs, results):
for o in outp:
o.op._inputs = list(inp)
graph.add_node(o)
for n in inp:
graph.add_node(n)
for o in outp:
graph.add_edge(n, o)
# mark initial assigns
fixed_assigns = dict()
op_states = dict()
for idx in range(2):
for i in range(2):
fixed_assigns[inputs[idx][i].op.key] = 'w%d' % (idx + 1)
op_states[results[idx][i].op.key] = OperandState.READY
fixed_assigns[results[idx][i].op.key] = 'w%d' % (idx + 1)
for inp in inputs:
for n in inp:
if n.op.key in fixed_assigns:
continue
op_states[n.op.key] = OperandState.READY
worker_metrics = dict(w1=24, w2=24, w3=24)
analyzer = GraphAnalyzer(graph, worker_metrics, fixed_assigns,
op_states)
assignments = analyzer.calc_operand_assignments(
analyzer.get_initial_operand_keys())
for inp in inputs:
if any(n.op.key in fixed_assigns for n in inp):
continue
self.assertEqual(1, len(set(assignments[n.op.key] for n in inp)))
worker_assigns = dict((k, 0) for k in worker_metrics)
for w in assignments.values():
worker_assigns[w] += 1
self.assertEqual(2, worker_assigns['w1'])
self.assertEqual(2, worker_assigns['w2'])
self.assertEqual(4, worker_assigns['w3'])
def test_fuse():
"""
test compose in build graph and optimize
"""
r"""
graph(@: node, #: composed_node):
@ --> @ --> @ ========> #
"""
chunks = [
TensorTreeAdd(args=[], _key=str(n)).new_chunk(None, None).data
for n in range(3)
]
graph = ChunkGraph([])
for c in chunks:
graph.add_node(c)
graph.add_edge(chunks[0], chunks[1])
graph.add_edge(chunks[1], chunks[2])
graph2 = graph.copy()
graph2._result_chunks = [chunks[2]]
_, fused_nodes = Fusion(graph2).fuse()
assert fused_nodes[0].composed == chunks[:3]
# make the middle one as result chunk, thus the graph cannot be composed
graph3 = graph.copy()
graph3._result_chunks = [chunks[1]]
_, fused_nodes = Fusion(graph3).fuse()
assert fused_nodes[0].composed == chunks[:2]
r"""
graph(@: node, #: composed_node):
@ @ @ @
\ / \ /
@ --> @ ========> #
/ \ / \
@ @ @ @
"""
chunks = [
TensorTreeAdd(args=[], _key=str(n)).new_chunk(None, None).data
for n in range(6)
]
graph = ChunkGraph([chunks[4], chunks[5]])
for c in chunks:
graph.add_node(c)
chunks[2].op._inputs = [chunks[0], chunks[1]]
chunks[3].op._inputs = [chunks[2]]
chunks[4].op._inputs = [chunks[3]]
chunks[5].op._inputs = [chunks[3]]
graph.add_edge(chunks[0], chunks[2])
graph.add_edge(chunks[1], chunks[2])
graph.add_edge(chunks[2], chunks[3])
graph.add_edge(chunks[3], chunks[4])
graph.add_edge(chunks[3], chunks[5])
_, fused_nodes = Fusion(graph).fuse()
assert fused_nodes[0].composed == chunks[2:4]
# to make sure the predecessors and successors of compose are right
# 0 and 1's successors must be composed
assert fused_nodes[0] in graph.successors(chunks[0])
assert fused_nodes[0] in graph.successors(chunks[1])
# check composed's inputs
assert chunks[0] in fused_nodes[0].inputs
assert chunks[1] in fused_nodes[0].inputs
# check composed's predecessors
assert chunks[0] in graph.predecessors(fused_nodes[0])
assert chunks[1] in graph.predecessors(fused_nodes[0])
# check 4 and 5's inputs
assert fused_nodes[0] in graph.successors(fused_nodes[0])[0].inputs
assert fused_nodes[0] in graph.successors(fused_nodes[0])[0].inputs
# check 4 and 5's predecessors
assert fused_nodes[0] in graph.predecessors(chunks[4])
assert fused_nodes[0] in graph.predecessors(chunks[5])
async def test_assign_cpu_tasks(actor_pool):
pool, session_id, assigner_ref, cluster_api, meta_api = actor_pool
input1 = TensorFetch(key='a', source_key='a',
dtype=np.dtype(int)).new_chunk([])
input2 = TensorFetch(key='b', source_key='b',
dtype=np.dtype(int)).new_chunk([])
input3 = TensorFetch(key='c', source_key='c',
dtype=np.dtype(int)).new_chunk([])
result_chunk = TensorTreeAdd(args=[input1, input2, input3]) \
.new_chunk([input1, input2, input3])
chunk_graph = ChunkGraph([result_chunk])
chunk_graph.add_node(input1)
chunk_graph.add_node(input2)
chunk_graph.add_node(input3)
chunk_graph.add_node(result_chunk)
chunk_graph.add_edge(input1, result_chunk)
chunk_graph.add_edge(input2, result_chunk)
chunk_graph.add_edge(input3, result_chunk)
await meta_api.set_chunk_meta(input1,
memory_size=200,
store_size=200,
bands=[('address0', 'numa-0')])
await meta_api.set_chunk_meta(input2,
memory_size=400,
store_size=400,
bands=[('address1', 'numa-0')])
await meta_api.set_chunk_meta(input3,
memory_size=400,
store_size=400,
bands=[('address2', 'numa-0')])
await cluster_api.set_node_status(node='address1',
role=NodeRole.WORKER,
status=NodeStatus.STOPPING)
await cluster_api.set_node_status(node='address3',
role=NodeRole.WORKER,
status=NodeStatus.STOPPING)
subtask = Subtask('test_task', session_id, chunk_graph=chunk_graph)
[result] = await assigner_ref.assign_subtasks([subtask])
assert result in (('address0', 'numa-0'), ('address2', 'numa-0'))
subtask.expect_bands = [('address0', 'numa-0')]
[result] = await assigner_ref.assign_subtasks([subtask])
assert result == ('address0', 'numa-0')
subtask.expect_bands = [('address0', 'numa-0'), ('address1', 'numa-0')]
[result] = await assigner_ref.assign_subtasks([subtask])
assert result == ('address0', 'numa-0')
subtask.expect_bands = [('address1', 'numa-0')]
[result] = await assigner_ref.assign_subtasks([subtask])
assert result in (('address0', 'numa-0'), ('address2', 'numa-0'))
result_chunk.op.gpu = True
subtask = Subtask('test_task', session_id, chunk_graph=chunk_graph)
with pytest.raises(NoMatchingSlots) as err:
await assigner_ref.assign_subtasks([subtask])
assert 'gpu' in str(err.value)
