def main(xls_file, start_cmd, end_cmd, cleanmr):
if start_cmd:
assert end_cmd is not None, "Wrong input"
if end_cmd:
assert start_cmd is not None, "Wrong input"
if cleanmr:
os.system("cleanmr")
if start_cmd:
os.system(start_cmd)
srtml.init()
df = pd.read_excel(xls_file, sheet_name="Model Raw Profile")
for index, row in df.iterrows():
pgraph = create_pgraph(df.loc[index, "Model Name"])
profile_dict = convert_profiles_to_regression_models(
json.loads(df.loc[index, "raw profile"]))
dinfo = DATASET_INFORMATION(
**json.loads(df.loc[index, "Dataset Information"]))
pgraph.submit(
df.loc[index, "Accuracy"],
dinfo,
df.loc[index, "feature"],
profile_dict,
)
shutdown()
if end_cmd:
os.system(end_cmd)
def main(xls_file, start_cmd, end_cmd):
if start_cmd:
os.system(start_cmd)
srtml.init()
df = pd.read_excel(xls_file, sheet_name="Arrival Information")
columns = [
"mu (qps)",
"cv",
"# requests",
"Latency Constraint (ms)",
"Planner",
"configuration",
"Estimated Latency (ms)",
"Cost",
"Accuracy",
]
new_df = pd.DataFrame(columns=columns)
for index, row in df.iterrows():
row_df = {columns[i]: df.loc[index, columns[i]] for i in range(5)}
with VGraph(name="classifier") as graph:
classifier = Classifier(
feature=IMAGE_CLASSIFICATION_FEATURE,
dataset_information=dinfo,
name="VClassifier",
)
arrival_curve = generate_fixed_arrival_process(
mean_qps=df.loc[index, columns[0]],
cv=df.loc[index, columns[1]],
num_requests=df.loc[index, columns[2]],
).tolist()
config = graph.configure(
throughput_qps_constraint=df.loc[index, columns[0]],
latency_ms_constraint=df.loc[index, columns[3]],
planner_kwargs={"inter_arrival_process": arrival_curve},
planner_cls=PLANNER_CLS.get(df.loc[index, columns[4]], None),
print_state=True,
materialize=False,
)
estimated_values = graph.state.get_estimate_values()._asdict()
row_df["configuration"] = json.dumps(config)
row_df["Estimated Latency (ms)"] = estimated_values["latency"]
row_df["Cost"] = estimated_values["cost"]
row_df["Accuracy"] = estimated_values["accuracy"]
new_df = new_df.append(row_df, ignore_index=True)
shutdown()
if end_cmd:
os.system(end_cmd)
with pd.ExcelWriter(xls_file, mode="a") as writer:
new_df.to_excel(writer, sheet_name="Planner Configuration")
def main(xls_file, start_cmd, end_cmd):
if start_cmd:
assert end_cmd is not None, "Wrong input"
if end_cmd:
assert start_cmd is not None, "Wrong input"
df = pd.read_excel(xls_file, sheet_name="Model Information")
clean_profile_df = pd.DataFrame()
raw_profile_df = pd.DataFrame(columns=[
"pgraph",
"raw profile",
"Dataset Information",
"feature",
"Model Name",
"Accuracy",
"sysinfo",
])
for index, row in df.iterrows():
if start_cmd:
os.system(start_cmd)
srtml.init()
# profile_df = pd.DataFrame(
# columns=[
# "pgraph"
# ]
# )
pgraph = create_pgraph(df.loc[index, "Model Name"])
profile_dict = profile_pgraph(
pgraph, **json.loads(df.loc[index, "profile configuration"]))
pprint(profile_dict)
clean_profile_df = pd.concat([
clean_profile_df,
get_dataframe_from_profile(pgraph.ppu_identifier, profile_dict),
])
raw_profile_df = raw_profile_df.append(
{
"pgraph": pgraph.ppu_identifier,
"raw profile": json.dumps(profile_dict),
"Dataset Information": df.loc[index, "Dataset Information"],
"feature": df.loc[index, "feature"],
"Model Name": df.loc[index, "Model Name"],
"Accuracy": df.loc[index, "Accuracy"],
"sysinfo": get_sysinfo(),
},
ignore_index=True,
)
shutdown()
if end_cmd:
os.system(end_cmd)
with pd.ExcelWriter(xls_file, mode="a") as writer:
clean_profile_df.to_excel(writer, sheet_name="Model Profile")
raw_profile_df.to_excel(writer, sheet_name="Model Raw Profile")
def main(xls_file, start_cmd, end_cmd, cleanmr):
if start_cmd:
assert end_cmd is not None, "Wrong input"
if end_cmd:
assert start_cmd is not None, "Wrong input"
if cleanmr:
os.system("cleanmr")
if start_cmd:
os.system(start_cmd)
ray_serve_kwargs = {
"ray_init_kwargs": {
"object_store_memory":
int(5e10),
"num_cpus":
24,
"_internal_config":
json.dumps({
"max_direct_call_object_size": 1000 * 1024 * 1024, # 10Mb
"max_grpc_message_size": 10000 * 1024 * 1024, # 100Mb
}),
# "resources": resources,
},
"start_server": False,
}
srtml.init()
df = pd.read_excel(xls_file, sheet_name="Model Raw Profile")
for index, row in df.iterrows():
pgraph = create_pgraph(df.loc[index, "Model Name"])
profile_dict = convert_profiles_to_regression_models(
json.loads(df.loc[index, "raw profile"]))
dinfo = DATASET_INFORMATION(
**json.loads(df.loc[index, "Dataset Information"]))
pgraph.submit(
df.loc[index, "Accuracy"],
dinfo,
df.loc[index, "feature"],
profile_dict,
)
shutdown()
if end_cmd:
os.system(end_cmd)
is_cuda=True,
)
# connection
# prepoc >> model_2
return graph
ray_serve_kwargs = {
"ray_init_kwargs": {
"object_store_memory":
int(5e10),
"num_cpus":
24,
"_internal_config":
json.dumps({
"max_direct_call_object_size": 10000 * 1024 * 1024, # 10Mb
"max_grpc_message_size": 100000 * 1024 * 1024, # 100Mb
}),
# "resources": resources,
},
"start_server": False,
}
srtml.init(ray_serve_kwargs=ray_serve_kwargs)
graph = create_pgraph()
graph.configure(SERVE_MODE)
graph.provision(SERVE_MODE)
s = {"warmup": 2, "num_requests": 5, "percentile": 99}
pprint(profile_pgraph(graph, **s))
