1 #! /usr/bin/env python3
3 # The goal is to introduce random failures in a simulation, to test simgrid under extreme conditions.
5 # It is made of several components.
7 # * a plugin: cmonkey. Can be used from the command line as follows:
8 # * --cfg=plugin:cmonkey --cfg=cmonkey/tell:1
9 # Get information about the resource count and the timestamps of each scheduling rounds.
10 # * --cfg=plugin:cmonkey --cfg=cmonkey/time:42 --cfg=cmonkey/host:1
11 # Kill the host #1 after 42 seconds (using a kernel::Timer so that no actor gets involved in the killing)
12 # * --cfg=plugin:cmonkey --cfg=cmonkey/time:42 --cfg=cmonkey/link:0
13 # Kill the link #0 after 42 seconds (using a kernel::Timer)
15 # * a python script: tools/simgrid-monkey (this file)
16 # * It takes a regular simgrid simulation as a parameter, use the cmonkey plugin to get the information about it,
17 # and then restart many runs, with one resource being turn_off() + turn_on() in each run.
18 # * Each resource gets killed between each timestamps, and on each timestamp.
19 # * So the amount of simulations is: 1 + (host_c+link_c) * timestamps * 2
21 # * Test program, written to resist these extreme conditions:
22 # * teshsuite/s4u/monkey-masterworkers: tests synchronous comms and execs (C++ and python)
23 # * teshsuite/s4u/monkey-semaphore: tests async semaphores (C++ only)
25 import multiprocessing as mp
35 cmd_tell = copy.deepcopy(cmd)
36 cmd_tell.append("--cfg=plugin:cmonkey")
37 cmd_tell.append("--cfg=cmonkey/tell:1")
38 cmd_tell.append("--log=root.t:critical")
39 cmd_tell.append("--log=cmonkey.t:info")
40 cmd_tell.append("--log=cmonkey.fmt:%m%n")
41 print(f"Get the initial info from the command.")
42 # print(f"from ``{' '.join(cmd_tell)}``")
43 first_run = subprocess.run(cmd_tell, shell=False, cwd=os.getcwd(), stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
45 if first_run.returncode != 0:
46 msg = f"Peek run of the command raised an error (retcode: {first_run.returncode})\n"
47 msg += f"Full command was {' '.join(cmd_tell)}\n"
48 if first_run.stdout is not None:
49 msg += str(first_run.stdout, errors='replace')
55 for line in str(first_run.stdout, errors='replace').split("\n"):
56 if re.match("^HOST_COUNT=(.*)", line):
57 m = re.match("^HOST_COUNT=(.*)", line)
58 host_count = int(m.group(1))
59 if re.match("^LINK_COUNT=(.*)", line):
60 m = re.match("^LINK_COUNT=(.*)", line)
61 link_count = int(m.group(1))
62 if re.match("^TIMESTAMP=(.*)", line):
63 m = re.match("^TIMESTAMP=(.*)", line)
64 timestamps.append(float(m.group(1)))
66 #print(f"hosts:{host_count} links:{link_count} timestamps:{' '.join(([str(i) for i in timestamps]))}")
67 return (host_count, link_count, timestamps)
69 parser = argparse.ArgumentParser(description='Run a simgrid simulation, and turn off/on resources at random.')
70 parser.add_argument('--valgrind', help="Run the simulations in valgrind")
71 parser.add_argument('command', nargs='*')
72 args = parser.parse_args()
74 (host_count, link_count, timestamps) = get_info(args.command)
75 timestamps = sorted([*{*timestamps}]) # kill duplicates
76 print(f"Monkey informations: hosts:{host_count} links:{link_count} timestamps:{' '.join(([str(i) for i in timestamps]))}")
80 def do_run(cmd, extra_params, test_todo):
81 global test_count, error_count
82 test_count = test_count + 1
83 cmd = copy.deepcopy(cmd)
84 cmd.append("--cfg=plugin:cmonkey")
85 for p in extra_params:
87 print(f"Start {' '.join(cmd)}")
90 run = subprocess.run(cmd, shell=False, cwd=os.getcwd(), stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
92 out = str(run.stdout, errors='replace')
93 if run.returncode != 0:
94 msg = f"ERROR (retcode: {run.returncode}). Output:\n"
98 for line in out.split("\n"):
99 if re.match("==.* in use at exit: ", line) and not re.match("==.* in use at exit: 0 bytes in 0 blocks", line):
100 m = re.match("==.* in use at exit: (.*)", line)
101 print(f"LEAK SUMMARY: {m.group(1)} in use at exit")
104 if re.match("==.* ERROR SUMMARY: ", line):
105 m = re.match("==.* ERROR SUMMARY: (.*)", line)
106 print(f"valgrind summary: {m.group(1)}")
107 if not re.match("==.* 0 errors from 0 contexts", line):
109 print (f"Test {test_count} out of {test_todo} succeded.\n")
115 test_todo = 2 * len(timestamps) * (host_count + link_count)
116 for pos in range(len(timestamps)):
117 now = timestamps[pos]
118 for host in range(host_count):
119 do_run(args.command, [f"--cfg=cmonkey/time:{(now-prev_time)/2}", f"--cfg=cmonkey/host:{host}"], test_todo)
120 for link in range(link_count):
121 do_run(args.command, [f"--cfg=cmonkey/time:{(now-prev_time)/2}", f"--cfg=cmonkey/link:{link}"], test_todo)
122 for host in range(host_count):
123 do_run(args.command, [f"--cfg=cmonkey/time:{now}", f"--cfg=cmonkey/host:{host}"], test_todo)
124 for link in range(link_count):
125 do_run(args.command, [f"--cfg=cmonkey/time:{now}", f"--cfg=cmonkey/link:{link}"], test_todo)
128 print(f"In total, the monkey found {error_count} errors.")
129 sys.exit(error_count)