def run_simrd(g, heuristic, budgets, liveness):
logger.info('Evaluating simrd ({}), liveness {}...'.format(
type(heuristic).__name__, 'enabled' if liveness else 'disabled'
))
futures = []
remote_simrd = ray.remote(num_cpus=NUM_ILP_CORES)(solve_simrd).remote
for b in budgets:
future = remote_simrd(
g, b, heuristic=heuristic, runtime=RuntimeV2EagerOptimized,
thrash=2.0, liveness=liveness
)
futures.append(future)
results = get_futures(futures, desc='simrd ({})'.format(type(heuristic).__name__))
return results
solve_checkpoint_last_node(g)
]
result_dict[SolveStrategy.CHEN_SQRTN_NOAP] = [solve_chen_sqrtn(g, False)]
result_dict[SolveStrategy.CHEN_SQRTN] = [solve_chen_sqrtn(g, True)]
# sweep chen's greedy baseline
logger.info("Running Chen's greedy baseline (APs only)")
chen_sqrtn_noap = result_dict[SolveStrategy.CHEN_SQRTN_NOAP][0]
greedy_eval_points = chen_sqrtn_noap.schedule_aux_data.activation_ram * (
1.0 + np.arange(-1, 2, 0.01))
remote_solve_chen_greedy = ray.remote(num_cpus=1)(solve_chen_greedy).remote
futures = [
remote_solve_chen_greedy(g, float(b), False)
for b in greedy_eval_points
]
result_dict[SolveStrategy.CHEN_GREEDY] = get_futures(
list(futures), desc="Greedy (APs only)")
if model_name not in CHAIN_GRAPH_MODELS:
logger.info(
"Running Chen's greedy baseline (no AP) as model is non-linear")
futures = [
remote_solve_chen_greedy(g, float(b), True)
for b in greedy_eval_points
]
result_dict[SolveStrategy.CHEN_SQRTN_NOAP] = get_futures(
list(futures), desc="Greedy (No AP)")
# sweep griewank baselines
logger.error(
"Skipping Griewank baselines as it was broken in parasj/checkmate#65")
# if model_name in CHAIN_GRAPH_MODELS:
# logger.info(f"Running Griewank baseline (APs only)")
remote_solve_chen_greedy(g, float(b), False)
for b in greedy_eval_points
])
futures.extend([
remote_solve_chen_greedy(g, float(b), True)
for b in greedy_eval_points
])
# # sweep griewank baselines
# if model_name in CHAIN_GRAPH_MODELS:
# solve_griewank(g, 1) # prefetch griewank solution from s3, otherwise ray will cause race condition
# griewank_eval_points = range(1, g.size + 1)
# remote_solve_griewank = ray.remote(num_cpus=1)(solve_griewank).remote
# futures.extend([remote_solve_griewank(g, float(b)) for b in griewank_eval_points])
for result in get_futures(futures, desc="Batch size: {}".format(bs)):
result_dict[bs][result.solve_strategy].append(result)
ray.shutdown()
max_batch_sizes = defaultdict(int)
for bs, strategy_results in result_dict.items():
for strategy, results in strategy_results.items():
is_valid = (
lambda r: r.schedule_aux_data is not None and r.
schedule_aux_data.peak_ram <= platform_ram - bs_param_ram_cost[
bs] and r.schedule_aux_data.cpu <= bs_fwd2xcost[bs])
if any(map(is_valid, results)):
max_batch_sizes[strategy] = max(bs, max_batch_sizes[strategy])
logging.info(
"SolveStrategy {} succeeded at batch size {}".format(
remote_solve_chen_greedy(g, float(b), False)
for b in greedy_eval_points
])
futures.extend([
remote_solve_chen_greedy(g, float(b), True)
for b in greedy_eval_points
])
# # sweep griewank baselines
# if model_name in CHAIN_GRAPH_MODELS:
# solve_griewank(g, 1) # prefetch griewank solution from s3, otherwise ray will cause race condition
# griewank_eval_points = range(1, g.size + 1)
# remote_solve_griewank = ray.remote(num_cpus=1)(solve_griewank).remote
# futures.extend([remote_solve_griewank(g, float(b)) for b in griewank_eval_points])
for result in get_futures(futures, desc=f"Batch size: {bs}"):
result_dict[bs][result.solve_strategy].append(result)
ray.shutdown()
max_batch_sizes = defaultdict(int)
for bs, strategy_results in result_dict.items():
for strategy, results in strategy_results.items():
is_valid = lambda r: r.schedule_aux_data is not None \
and r.schedule_aux_data.peak_ram <= platform_ram - bs_param_ram_cost[bs] \
and r.schedule_aux_data.cpu <= bs_fwd2xcost[bs]
if any(map(is_valid, results)):
max_batch_sizes[strategy] = max(bs,
max_batch_sizes[strategy])
logging.info(
f"SolveStrategy {strategy} succeeded at batch size {bs}"
solve_checkpoint_last_node(g)
]
result_dict[SolveStrategy.CHEN_SQRTN_NOAP] = [solve_chen_sqrtn(g, False)]
result_dict[SolveStrategy.CHEN_SQRTN] = [solve_chen_sqrtn(g, True)]
# sweep chen's greedy baseline
logger.info(f"Running Chen's greedy baseline (No AP)")
chen_sqrtn_noap = result_dict[SolveStrategy.CHEN_SQRTN_NOAP][0]
greedy_eval_points = chen_sqrtn_noap.schedule_aux_data.activation_ram * (
1. + np.arange(-1, 2, 0.01))
remote_solve_chen_greedy = ray.remote(num_cpus=1)(solve_chen_greedy).remote
futures = [
remote_solve_chen_greedy(g, float(b), False)
for b in greedy_eval_points
]
result_dict[SolveStrategy.CHEN_GREEDY_NOAP] = get_futures(
list(futures), desc="Greedy (No AP)")
if model_name not in CHAIN_GRAPH_MODELS:
logger.info(
f"Running Chen's greedy baseline (AP) as model is non-linear")
futures = [
remote_solve_chen_greedy(g, float(b), True)
for b in greedy_eval_points
]
result_dict[SolveStrategy.CHEN_GREEDY] = get_futures(
list(futures), desc="Greedy (APs only)")
# sweep griewank baselines
if model_name in CHAIN_GRAPH_MODELS:
logger.info(f"Running Griewank baseline (APs only)")
clean_griewank_cache()
solve_griewank(