def test_lifo_scheduler_favor_cpu_for_requests_without_accelerators(
example_mixed_agents,
example_pending_sessions,
):
# Check the reverse with the LIFO scheduler.
# The result must be same.
scheduler = LIFOSlotScheduler({})
for idx in range(3):
picked_session_id = scheduler.pick_session(example_total_capacity,
example_pending_sessions,
[])
assert picked_session_id == example_pending_sessions[-1].session_id
picked_session = _find_and_pop_picked_session(example_pending_sessions,
picked_session_id)
agent_id = scheduler.assign_agent_for_session(example_mixed_agents,
picked_session)
if idx == 2:
# example_mixed_agents do not have any agent with ROCM accelerators.
assert agent_id is None
elif idx == 1:
assert agent_id == AgentId('i-gpu')
elif idx == 0:
# It should favor the CPU-only agent if the requested slots
# do not include accelerators.
assert agent_id == AgentId('i-cpu')
def example_mixed_agents():
return [
AgentContext(
agent_id=AgentId('i-gpu'),
agent_addr='10.0.1.1:6001',
scaling_group='sg01',
available_slots=ResourceSlot({
'cpu': Decimal('4.0'),
'mem': Decimal('4096'),
'cuda.shares': Decimal('4.0'),
}),
occupied_slots=ResourceSlot({
'cpu': Decimal('0'),
'mem': Decimal('0'),
'cuda.shares': Decimal('0'),
}),
),
AgentContext(
agent_id=AgentId('i-cpu'),
agent_addr='10.0.2.1:6001',
scaling_group='sg02',
available_slots=ResourceSlot({
'cpu': Decimal('3.0'),
'mem': Decimal('2560'),
'cuda.shares': Decimal('0'),
}),
occupied_slots=ResourceSlot({
'cpu': Decimal('0'),
'mem': Decimal('0'),
'cuda.shares': Decimal('0'),
}),
),
]
def test_fifo_scheduler(example_agents, example_pending_sessions,
example_existing_sessions):
scheduler = FIFOSlotScheduler({})
picked_session_id = scheduler.pick_session(example_total_capacity,
example_pending_sessions,
example_existing_sessions)
assert picked_session_id == example_pending_sessions[0].kernel_id
picked_session = _find_and_pop_picked_session(example_pending_sessions,
picked_session_id)
agent_id = scheduler.assign_agent(example_agents, picked_session)
assert agent_id == AgentId('i-001')
def example_agents_no_valid():
return [
AgentContext(
agent_id=AgentId('i-001'),
agent_addr='10.0.1.1:6001',
scaling_group='sg01',
available_slots=ResourceSlot({
'cpu': Decimal('0'),
'mem': Decimal('0'),
'cuda.shares': Decimal('0'),
'rocm.devices': Decimal('0'),
}),
occupied_slots=ResourceSlot({
'cpu': Decimal('4.0'),
'mem': Decimal('4096'),
'cuda.shares': Decimal('4.0'),
'rocm.devices': Decimal('2'),
}),
),
AgentContext(
agent_id=AgentId('i-101'),
agent_addr='10.0.2.1:6001',
scaling_group='sg02',
available_slots=ResourceSlot({
'cpu': Decimal('0'),
'mem': Decimal('0'),
'cuda.shares': Decimal('0'),
'rocm.devices': Decimal('0'),
}),
occupied_slots=ResourceSlot({
'cpu': Decimal('3.0'),
'mem': Decimal('2560'),
'cuda.shares': Decimal('1.0'),
'rocm.devices': Decimal('8'),
}),
),
]
def scb(app_ctx: web.Application, agent_id: AgentId, event_name: str):
assert app_ctx is app
assert agent_id == AgentId('i-test')
assert event_name == 'test-event-02'
raise OverflowError
async def acb(app_ctx: web.Application, agent_id: AgentId,
event_name: str):
assert app_ctx is app
assert agent_id == AgentId('i-test')
assert event_name == 'test-event-02'
raise ZeroDivisionError
def scb(app_ctx: web.Application, agent_id: AgentId, event_name: str):
assert app_ctx is app
assert agent_id == AgentId('i-test')
assert event_name == 'test-event-01'
records['test-var'].add('sync')