forked from boringlee24/socc22-miso
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcontroller_helper.py
167 lines (153 loc) · 6.9 KB
/
controller_helper.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
import pdb
import time
import os
user = os.environ.get('USER')
import random
import json
import numpy as np
import glob
import argparse
import math
from pathlib import Path
import sys
sys.path.append(f'/home/{user}/GIT/socc22-miso/mps/scheduler/simulator/')
from utils import *
import copy
sys.path.append(f'/home/{user}/GIT/socc22-miso/workloads/')
from send_signal import send_signal
import socket
import threading
import _thread
import signal
from tcp_interpreter import *
from threading import Event
# start job
def start_job(node, job, gpu, sliceid):
cmd = f'start {job} gpu {gpu} slice {sliceid}'
send_signal(node, cmd=cmd)
def mps_start(node, job, gpu, level=50):
cmd = f'mps_strt {job} gpu {gpu} lvl {level}'
send_signal(node, cmd=cmd)
def mps_resume(node, job, gpu, resume_batch, level=50):
cmd = f'mps_rsm {job} gpu {gpu} batch {resume_batch} lvl {level}'
send_signal(node, cmd=cmd)
def resume_job(node, job, gpu, sliceid, resume_batch):
cmd = f'resume {job} gpu {gpu} slice {sliceid} batch {resume_batch}'
send_signal(node, cmd=cmd)
def config_gpu(node, gpu, partition):
cmd = f'config gpu {gpu} partition {partition}'
send_signal(node, cmd=cmd)
def start_mps(node, gpu): # this is essentially reset_mig and create_ins 7g.40gb, then run ./enable_mps_on_mig
cmd = f'mps_enable {gpu}'
send_signal(node, cmd=cmd)
def end_mps(node, gpu):
cmd = f'mps_disable {gpu}'
send_signal(node, cmd=cmd)
def fkill_job(node, job, pid):
cmd = f'fkill {job} pid {pid}'
send_signal(node, cmd=cmd)
def kill_all(node):
cmd = 'kill all'
send_signal(node, cmd=cmd)
def broadcast_host(node, runtime):
cmd = f'hostname {socket.gethostname()}'
send_signal(node, cmd=cmd)
cmd = f'log_dir {runtime.tc}'
send_signal(node, cmd=cmd)
def save_jobs(node, job_list, runtime, run_log):
finish_status = [runtime.finish[job] for job in job_list]
if 1 in finish_status:
print(f'checkpoing {job_list} is invalid due to finish status {finish_status}', file=run_log, flush=True) #TODO: see if something can be done for jobs that are near finishing
return False # indicate this save is invalid
for job in job_list:
if runtime.pid_dict[job] == 0:
raise RuntimeError(f'job {job} PID is not updated')
cmd = f'save {job} pid {runtime.pid_dict[job]}'
send_signal(node, cmd=cmd)
# wait till all ckpt signals are received
print(f'waiting for checkpoint to finish, jobs {str(job_list)}', file=run_log, flush=True)
while True:
# make sure all ckpt_dict are 1
time.sleep(0.5) #TODO: 1 pdb.set_trace()
ckpt_sum = 0
for job in job_list:
ckpt_sum += runtime.ckpt_dict[job]
if ckpt_sum == len(job_list):
print('checkpoint finished', file=run_log, flush=True)
for job in job_list:
fkill_job(node, job, runtime.pid_dict[job])
return True
def thread_func(event, runtime, run_log, mode='full'): # this is an instance of the Experiment class
# here listen on the socket
sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
server_address = (socket.gethostname(), 10002)
print('starting up on {} port {}'.format(*server_address), file=run_log, flush=True)
sock.bind(server_address)
sock.listen(5)
while not event.is_set():
# Wait for a connection
connection, client_address = sock.accept()
try:
while not event.is_set():
data = connection.recv(32)
if data:
data_str = data.decode('utf-8')
if 'term_thread' in data_str:
connection.sendall(b'success')
print('breaking connection loop')
sys.exit()
# connection.close()
elif mode == 'full':
interpret_full(data_str, runtime, run_log)
elif mode == 'mps':
interpret_mps(data_str, runtime, run_log)
elif mode == 'static':
interpret_static(data_str, runtime, run_log)
elif mode == 'oracle' or mode == 'miso':
interpret_miso(data_str, runtime, run_log)
# elif 'waste' in data_str:
# global epoch_waste_dict
# job_name = data_str.split(' ')[0]
# epoch_waste_time = data_str.split(' ')[2]
# epoch_waste_dict[job_name] += int(epoch_waste_time)
# elif 'b_end' in data_str:
# job_name = data_str.split(' ')[0]
# job = job_name.replace('job','')
# ovhd_b[job].append(int(time.time() - b_start[job]))
# c_start[job] = time.time()
# elif 'c_end' in data_str:
# job_name = data_str.split(' ')[0]
# job = job_name.replace('job','')
# ovhd_c[job].append(int(time.time() - c_start[job]))
# d_start[job] = time.time()
# elif 'd_end' in data_str:
# job_name = data_str.split(' ')[0]
# job = job_name.replace('job','')
# ovhd_d[job].append(int(time.time() - d_start[job]))
# ovhd_total[job].append(int(time.time() - ovhd_start[job]))
# if ovhd_start[job] != 0:
# overhead[job] += int(time.time() - ovhd_start[job])
# ovhd_start[job] = 0
# if job in list(K80_job.values()):
# K80_start_time[job] = time.time()
# elif job in list(V100_job.values()):
# V100_start_time[job] = time.time()
# promote_start_time[job] = time.time()
# elif '1st_epoch' in data_str: # 'job50 1st_epoch 35'
# job_name = data_str.split(' ')[0]
# job = job_name.replace('job','')
# epoch_time = int(data_str.split(' ')[2])
# if job in list(K80_job.values()):
# k80_1st[job].append(epoch_time)
# elif job in list(V100_job.values()):
# v100_1st[job].append(epoch_time)
#if 'ckpt_qual' in data_str or 'finish' in data_str or 'checkpoint' in data_str:
# print('received ' + data_str)
connection.sendall(b'success')
#time.sleep(5)
else:
break
finally:
connection.close()
# print('Terminated connection')
print('Terminated thread')