def _test_peer_communication_tag(queue, rank, msg_size):
ctx = ucx_api.UCXContext(feature_flags=(ucx_api.Feature.TAG, ))
worker = ucx_api.UCXWorker(ctx)
queue.put((rank, worker.get_address()))
right_rank, right_address = queue.get()
left_rank, left_address = queue.get()
right_ep = ucx_api.UCXEndpoint.create_from_worker_address(
worker,
right_address,
endpoint_error_handling=True,
)
left_ep = ucx_api.UCXEndpoint.create_from_worker_address(
worker,
left_address,
endpoint_error_handling=True,
)
recv_msg = bytearray(msg_size)
if rank == 0:
send_msg = bytes(os.urandom(msg_size))
blocking_send(worker, right_ep, send_msg, right_rank)
blocking_recv(worker, left_ep, recv_msg, rank)
assert send_msg == recv_msg
else:
blocking_recv(worker, left_ep, recv_msg, rank)
blocking_send(worker, right_ep, recv_msg, right_rank)
def _server_probe(queue, transfer_api):
"""Server that probes and receives message after client disconnected.
Note that since it is illegal to call progress() in callback functions,
we keep a reference to the endpoint after the listener callback has
terminated, this way we can progress even after Python blocking calls.
"""
feature_flags = (ucx_api.Feature.AM
if transfer_api == "am" else ucx_api.Feature.TAG, )
ctx = ucx_api.UCXContext(feature_flags=feature_flags)
worker = ucx_api.UCXWorker(ctx)
# Keep endpoint to be used from outside the listener callback
ep = [None]
def _listener_handler(conn_request):
ep[0] = ucx_api.UCXEndpoint.create_from_conn_request(
worker,
conn_request,
endpoint_error_handling=True,
)
listener = ucx_api.UCXListener(worker=worker,
port=0,
cb_func=_listener_handler)
queue.put(listener.port),
while ep[0] is None:
worker.progress()
ep = ep[0]
# Ensure wireup and inform client before it can disconnect
if transfer_api == "am":
wireup = blocking_am_recv(worker, ep)
else:
wireup = bytearray(len(WireupMessage))
blocking_recv(worker, ep, wireup)
queue.put("wireup completed")
# Ensure client has disconnected -- endpoint is not alive anymore
while ep.is_alive() is True:
worker.progress()
# Probe/receive message even after the remote endpoint has disconnected
if transfer_api == "am":
while ep.am_probe() is False:
worker.progress()
received = blocking_am_recv(worker, ep)
else:
while worker.tag_probe(0) is False:
worker.progress()
received = bytearray(len(DataMessage))
blocking_recv(worker, ep, received)
assert wireup == WireupMessage
assert received == DataMessage
def _echo_client(msg_size, port, endpoint_error_handling):
ctx = ucx_api.UCXContext(feature_flags=(ucx_api.Feature.TAG,))
worker = ucx_api.UCXWorker(ctx)
ep = ucx_api.UCXEndpoint.create(
worker, "localhost", port, endpoint_error_handling=endpoint_error_handling,
)
send_msg = bytes(os.urandom(msg_size))
recv_msg = bytearray(msg_size)
blocking_send(worker, ep, send_msg)
blocking_recv(worker, ep, recv_msg)
assert send_msg == recv_msg
def _test_peer_communication_rma(queue, rank, msg_size):
ctx = ucx_api.UCXContext(feature_flags=(ucx_api.Feature.RMA,
ucx_api.Feature.TAG))
worker = ucx_api.UCXWorker(ctx)
self_address = worker.get_address()
mem_handle = ctx.alloc(msg_size)
self_base = mem_handle.address
self_prkey = mem_handle.pack_rkey()
self_ep, self_mem = _rma_setup(worker, self_address, self_prkey, self_base,
msg_size)
send_msg = bytes(repeat(rank, msg_size))
if not self_mem.put_nbi(send_msg):
blocking_flush(self_ep)
queue.put((rank, self_address, self_prkey, self_base))
right_rank, right_address, right_prkey, right_base = queue.get()
left_rank, left_address, left_prkey, left_base = queue.get()
right_ep, right_mem = _rma_setup(worker, right_address, right_prkey,
right_base, msg_size)
right_msg = bytearray(msg_size)
right_mem.get_nbi(right_msg)
left_ep, left_mem = _rma_setup(worker, left_address, left_prkey, left_base,
msg_size)
left_msg = bytearray(msg_size)
left_mem.get_nbi(left_msg)
blocking_flush(worker)
assert left_msg == bytes(repeat(left_rank, msg_size))
assert right_msg == bytes(repeat(right_rank, msg_size))
# We use the blocking tag send/recv as a barrier implementation
recv_msg = bytearray(8)
if rank == 0:
send_msg = bytes(os.urandom(8))
blocking_send(worker, right_ep, send_msg, right_rank)
blocking_recv(worker, left_ep, recv_msg, rank)
else:
blocking_recv(worker, left_ep, recv_msg, rank)
blocking_send(worker, right_ep, recv_msg, right_rank)
def client(queue, port, server_address, args):
if args.client_cpu_affinity >= 0:
os.sched_setaffinity(0, [args.client_cpu_affinity])
import numpy as np
if args.object_type == "numpy":
import numpy as xp
elif args.object_type == "cupy":
import cupy as xp
xp.cuda.runtime.setDevice(args.client_dev)
else:
import cupy as xp
import rmm
rmm.reinitialize(
pool_allocator=True,
managed_memory=False,
initial_pool_size=args.rmm_init_pool_size,
devices=[args.client_dev],
)
xp.cuda.runtime.setDevice(args.client_dev)
xp.cuda.set_allocator(rmm.rmm_cupy_allocator)
ctx = ucx_api.UCXContext(
feature_flags=(
ucx_api.Feature.AM if args.enable_am is True else ucx_api.Feature.TAG,
)
)
worker = ucx_api.UCXWorker(ctx)
register_am_allocators(args, worker)
ep = ucx_api.UCXEndpoint.create(
worker,
server_address,
port,
endpoint_error_handling=True,
)
send_msg = xp.arange(args.n_bytes, dtype="u1")
if args.reuse_alloc:
recv_msg = xp.zeros(args.n_bytes, dtype="u1")
if args.enable_am:
blocking_am_send(worker, ep, send_msg)
blocking_am_recv(worker, ep)
else:
wireup_recv = bytearray(len(WireupMessage))
blocking_send(worker, ep, WireupMessage)
blocking_recv(worker, ep, wireup_recv)
op_lock = Lock()
finished = [0]
outstanding = [0]
def maybe_progress():
while outstanding[0] >= args.max_outstanding:
worker.progress()
def op_started():
with op_lock:
outstanding[0] += 1
def op_completed():
with op_lock:
outstanding[0] -= 1
finished[0] += 1
if args.cuda_profile:
xp.cuda.profiler.start()
times = []
for i in range(args.n_iter + args.n_warmup_iter):
start = clock()
if args.enable_am:
blocking_am_send(worker, ep, send_msg)
blocking_am_recv(worker, ep)
else:
if not args.reuse_alloc:
recv_msg = xp.zeros(args.n_bytes, dtype="u1")
if args.delay_progress:
maybe_progress()
non_blocking_send(worker, ep, send_msg, op_started, op_completed)
maybe_progress()
non_blocking_recv(worker, ep, recv_msg, op_started, op_completed)
else:
blocking_send(worker, ep, send_msg)
blocking_recv(worker, ep, recv_msg)
stop = clock()
if i >= args.n_warmup_iter:
times.append(stop - start)
if args.delay_progress:
while finished[0] != 2 * (args.n_iter + args.n_warmup_iter):
worker.progress()
if args.cuda_profile:
xp.cuda.profiler.stop()
assert len(times) == args.n_iter
bw_avg = format_bytes(2 * args.n_iter * args.n_bytes / sum(times))
bw_med = format_bytes(2 * args.n_bytes / np.median(times))
lat_avg = int(sum(times) * 1e9 / (2 * args.n_iter))
lat_med = int(np.median(times) * 1e9 / 2)
delay_progress_str = (
f"True ({args.max_outstanding})" if args.delay_progress is True else "False"
)
print("Roundtrip benchmark")
print_separator(separator="=")
print_key_value(key="Iterations", value=f"{args.n_iter}")
print_key_value(key="Bytes", value=f"{format_bytes(args.n_bytes)}")
print_key_value(key="Object type", value=f"{args.object_type}")
print_key_value(key="Reuse allocation", value=f"{args.reuse_alloc}")
print_key_value(key="Transfer API", value=f"{'AM' if args.enable_am else 'TAG'}")
print_key_value(key="Delay progress", value=f"{delay_progress_str}")
print_key_value(key="UCX_TLS", value=f"{ucp.get_config()['TLS']}")
print_key_value(key="UCX_NET_DEVICES", value=f"{ucp.get_config()['NET_DEVICES']}")
print_separator(separator="=")
if args.object_type == "numpy":
print_key_value(key="Device(s)", value="CPU-only")
s_aff = (
args.server_cpu_affinity
if args.server_cpu_affinity >= 0
else "affinity not set"
)
c_aff = (
args.client_cpu_affinity
if args.client_cpu_affinity >= 0
else "affinity not set"
)
print_key_value(key="Server CPU", value=f"{s_aff}")
print_key_value(key="Client CPU", value=f"{c_aff}")
else:
print_key_value(key="Device(s)", value=f"{args.server_dev}, {args.client_dev}")
print_separator(separator="=")
print_key_value("Bandwidth (average)", value=f"{bw_avg}/s")
print_key_value("Bandwidth (median)", value=f"{bw_med}/s")
print_key_value("Latency (average)", value=f"{lat_avg} ns")
print_key_value("Latency (median)", value=f"{lat_med} ns")
if not args.no_detailed_report:
print_separator(separator="=")
print_key_value(key="Iterations", value="Bandwidth, Latency")
print_separator(separator="-")
for i, t in enumerate(times):
ts = format_bytes(2 * args.n_bytes / t)
lat = int(t * 1e9 / 2)
print_key_value(key=i, value=f"{ts}/s, {lat}ns")
def client(queue, port, server_address, args):
if args.client_cpu_affinity >= 0:
os.sched_setaffinity(0, [args.client_cpu_affinity])
import numpy as np
if args.object_type == "numpy":
import numpy as xp
elif args.object_type == "cupy":
import cupy as xp
xp.cuda.runtime.setDevice(args.client_dev)
else:
import cupy as xp
import rmm
rmm.reinitialize(
pool_allocator=True,
managed_memory=False,
initial_pool_size=args.rmm_init_pool_size,
devices=[args.client_dev],
)
xp.cuda.runtime.setDevice(args.client_dev)
xp.cuda.set_allocator(rmm.rmm_cupy_allocator)
ctx = ucx_api.UCXContext(feature_flags=(
ucx_api.Feature.AM if args.enable_am is True else ucx_api.Feature.TAG,
))
worker = ucx_api.UCXWorker(ctx)
register_am_allocators(args, worker)
ep = ucx_api.UCXEndpoint.create(
worker,
server_address,
port,
endpoint_error_handling=ucx_api.get_ucx_version() >= (1, 10, 0),
)
send_msg = xp.arange(args.n_bytes, dtype="u1")
if args.reuse_alloc:
recv_msg = xp.zeros(args.n_bytes, dtype="u1")
if args.enable_am:
blocking_am_send(worker, ep, send_msg)
blocking_am_recv(worker, ep)
else:
wireup_recv = bytearray(len(WireupMessage))
blocking_send(worker, ep, WireupMessage)
blocking_recv(worker, ep, wireup_recv)
op_lock = Lock()
finished = [0]
outstanding = [0]
def maybe_progress():
while outstanding[0] >= args.max_outstanding:
worker.progress()
def op_started():
with op_lock:
outstanding[0] += 1
def op_completed():
with op_lock:
outstanding[0] -= 1
finished[0] += 1
if args.cuda_profile:
xp.cuda.profiler.start()
times = []
for i in range(args.n_iter):
start = clock()
if args.enable_am:
blocking_am_send(worker, ep, send_msg)
blocking_am_recv(worker, ep)
else:
if not args.reuse_alloc:
recv_msg = xp.zeros(args.n_bytes, dtype="u1")
if args.delay_progress:
maybe_progress()
non_blocking_send(worker, ep, send_msg, op_started,
op_completed)
maybe_progress()
non_blocking_recv(worker, ep, recv_msg, op_started,
op_completed)
else:
blocking_send(worker, ep, send_msg)
blocking_recv(worker, ep, recv_msg)
stop = clock()
times.append(stop - start)
if args.delay_progress:
while finished[0] != 2 * args.n_iter:
worker.progress()
if args.cuda_profile:
xp.cuda.profiler.stop()
assert len(times) == args.n_iter
delay_progress_str = (f"True ({args.max_outstanding})"
if args.delay_progress is True else "False")
print("Roundtrip benchmark")
print("--------------------------")
print(f"n_iter | {args.n_iter}")
print(f"n_bytes | {format_bytes(args.n_bytes)}")
print(f"object | {args.object_type}")
print(f"reuse alloc | {args.reuse_alloc}")
print(f"transfer API | {'AM' if args.enable_am else 'TAG'}")
print(f"delay progress | {delay_progress_str}")
print(f"UCX_TLS | {ucp.get_config()['TLS']}")
print(f"UCX_NET_DEVICES | {ucp.get_config()['NET_DEVICES']}")
print("==========================")
if args.object_type == "numpy":
print("Device(s) | CPU-only")
s_aff = (args.server_cpu_affinity
if args.server_cpu_affinity >= 0 else "affinity not set")
c_aff = (args.client_cpu_affinity
if args.client_cpu_affinity >= 0 else "affinity not set")
print(f"Server CPU | {s_aff}")
print(f"Client CPU | {c_aff}")
else:
print(f"Device(s) | {args.server_dev}, {args.client_dev}")
avg = format_bytes(2 * args.n_iter * args.n_bytes / sum(times))
med = format_bytes(2 * args.n_bytes / np.median(times))
print(f"Average | {avg}/s")
print(f"Median | {med}/s")
if not args.no_detailed_report:
print("--------------------------")
print("Iterations")
print("--------------------------")
for i, t in enumerate(times):
ts = format_bytes(2 * args.n_bytes / t)
ts = (" " * (9 - len(ts))) + ts
print("%03d |%s/s" % (i, ts))