def save_data(self, start_end_times):
"""
Calculates the actual request rate and prints as the first line. Saves
the CDFs of latency and bandwidth betweeen 1-2 minutes to disk. Assumes
that all redis client hosts report correct times.
"""
# Save the CDF of the start times. TODO: Debug.
with open('data/distr_redis_raw_start_time_cdf.txt', 'w') as f:
start_time_list = [t for (t, _, _) in start_end_times]
for (t, p) in util.make_cdf_table(start_time_list):
print >> f, '%.5f' % t, p
# Save the start end times list. TODO: Debug.
with open('data/distr_redis_raw_start_end_times.txt','w') as f:
for (start_time, end_time, data_file) in start_end_times:
print >> f, '%.5f' % start_time, end_time, data_file
# Filter out irrelevant time values. Focus on 60th-120th seconds.
min_time = min([start_time for (start_time, _, _) in start_end_times])
def is_steady_state(start_end_time_tuple):
(start_time, _, _) = start_end_time_tuple
return min_time + 60 <= start_time <= min_time + 120
filtered_times = filter(is_steady_state, start_end_times)
filtered_times.sort()
print 'Raw data size:', len(start_end_times),
print 'Data between 60-120th seconds:', len(filtered_times)
# Figure out the actual gaps in milliseconds.
start_time_list = [start for (start, _, _) in filtered_times]
gap_list = []
for index in range(0, len(start_time_list) - 1):
gap = start_time_list[index + 1] - start_time_list[index]
gap_list.append(gap * 1000.0)
gap_list.sort()
print 'Client gap: (mean, stdev) =', util.get_mean_and_stdev(gap_list),
print 'median =', gap_list[len(gap_list)/2]
# Calculate latency and bandwidth.
latency_list = []
bandwidth_list = []
for (start_time, end_time, _) in filtered_times:
if end_time is None:
latency = -1
bandwidth = 0
else:
latency = end_time - start_time # seconds
bandwidth = DATA_LENGTH / latency # Bytes/s
latency_list.append(latency * 1000.0) # milliseconds
bandwidth_list.append(bandwidth * 8.0 / 1000000.0) # Mbps
# Write to file.
with open('data/distr_redis_latency.txt', 'w') as f:
for (v, p) in util.make_cdf_table(latency_list):
print >> f, v, p
with open('data/distr_redis_bw.txt', 'w') as f:
for (v, p) in util.make_cdf_table(bandwidth_list):
print >> f, v, p
def data_analysis():
start_end_times = []
for path in os.listdir('.'):
if path.startswith('simplified-') and path.endswith('.tmp'):
with open(path) as f:
start_end_times += pickle.load(f)
print 'Loaded', path
# Extract steady state.
min_time = min([start_time for (start_time, _) in start_end_times])
def is_steady_state(start_end_time_tuple):
(start_time, _) = start_end_time_tuple
return min_time + INTERESTING_TIME_START <= start_time <= min_time + INTERESTING_TIME_END
filtered_times = filter(is_steady_state, start_end_times)
filtered_times.sort()
print 'Raw data size:', len(start_end_times),
print 'Data at steady state:', len(filtered_times)
# Figure out the actual gaps in milliseconds.
start_time_list = [start for (start, _) in filtered_times]
gap_list = []
for index in range(0, len(start_time_list) - 1):
gap = start_time_list[index + 1] - start_time_list[index]
gap_list.append(gap * 1000.0)
gap_list.sort()
print 'Client gap: (mean, stdev) =', util.get_mean_and_stdev(gap_list),
print 'median =', gap_list[len(gap_list) / 2]
# Calculate latency and bandwidth.
latency_list = []
bandwidth_list = []
for (start_time, end_time) in filtered_times:
if end_time is None:
latency = -1
bandwidth = 0
else:
latency = end_time - start_time # seconds
bandwidth = DATA_LENGTH / latency # Bytes/s
latency_list.append(latency * 1000.0) # milliseconds
bandwidth_list.append(bandwidth * 8.0 / 1000000.0) # Mbps
# Write to file.
with open('data/simplified_redis_latency.csv', 'w') as f:
for (v, p) in util.make_cdf_table(latency_list):
print >> f, '%.10f,%.10f' % (v, p)
with open('data/simplified_redis_bw.csv', 'w') as f:
for (v, p) in util.make_cdf_table(bandwidth_list):
print >> f, '%.10f,%.10f' % (v, p)
def start_processes(process_count, worker_thread_per_process,
client_count, gap_ms, data_length, redis_host):
switch = Switch(config.active_config)
data_length = int(data_length)
total_workers = process_count * worker_thread_per_process
redis_set(data_length)
worker_status_queue = Queue(maxsize=total_workers)
client_id_queue = Queue(maxsize=client_count)
result_queue = Queue(maxsize=client_count)
# Starts the worker processes that spawn individual worker threads.
for _ in range(process_count):
p = Process(target=RedisClientProcess,
args=(worker_thread_per_process, data_length, redis_host,
worker_status_queue, client_id_queue, result_queue))
p.daemon = True
p.start()
# Wait for all worker threads to start.
while True:
started_count = worker_status_queue.qsize()
if started_count < total_workers:
print total_workers - started_count, 'workers yet to start.'
time.sleep(1)
else:
break
# Send requests in a different thread.
util.ping_test(dest_host=redis_host, how_many_pings=2)
def requests():
for client_id in range(client_count):
client_id_queue.put(client_id)
time.sleep(gap_ms / 1000.0)
t = threading.Thread(target=requests)
t.daemon = True
t.start()
# Monitor the changes for the first minute.
base_time = time.time()
while True:
current_count = result_queue.qsize()
remaining_count = client_count - current_count
print 'Current:', current_count, 'Remaining:', remaining_count
if remaining_count > 0 and time.time() - base_time < 120:
try:
time.sleep(10)
except KeyboardInterrupt:
break
if redis_host == REDIS_HOST_OF:
rule_list = switch.dump_tables(filter_str='')
print 't =', time.time() - base_time,
print '; tcam_size =', len([rule for rule in rule_list if 'table_id=0' in rule]),
print '; table_1_size =', len([rule for rule in rule_list if 'table_id=1' in rule]),
print '; table_2_size =', len([rule for rule in rule_list if 'table_id=2' in rule]),
print '; total_size =', len([rule for rule in rule_list if 'cookie' in rule])
else:
break
# Extract the result into local lists. All time values are expressed in ms.
# We're only interested in results between 30-60 seconds.
print 'Analyzing the result...'
start_time_list = []
completion_time_list = []
while not result_queue.empty():
(_, start_time, end_time) = result_queue.get()
if start_time - base_time >= 60:
start_time_list.append(start_time * 1000.0)
if end_time is None:
completion_time = -100.0 # Not to be plotted.
else:
completion_time = (end_time - start_time) * 1000.0
completion_time_list.append(completion_time)
# Calculate the actual request gap.
start_time_list.sort()
gap_list = []
for index in range(0, len(start_time_list) - 1):
gap_list.append(start_time_list[index + 1] - start_time_list[index])
print 'Client gap: (mean, stdev) =', util.get_mean_and_stdev(gap_list)
# Calculate the CDF of completion times.
cdf_list = util.make_cdf(completion_time_list)
with open('data/realistic_redis_completion_times.txt', 'w') as f:
for (x, y) in zip(completion_time_list, cdf_list):
print >> f, x, y
def data_analysis():
start_end_times = []
# Load experiment data file
try:
data_file = 'async-' + os.environ['EXP_NAME']
except KeyError:
data_file = None
for path in os.listdir('data'):
if (path == data_file) or \
(data_file is None and path.startswith('async-') and path.endswith('.tmp')):
with open('data/' + path) as f:
start_end_times += pickle.load(f)
print 'Loaded', path
# Extract steady state.
min_time = min([start_time for (start_time, _) in start_end_times])
def is_steady_state(start_end_time_tuple):
(start_time, _) = start_end_time_tuple
return min_time + INTERESTING_TIME_START <= start_time <= min_time + INTERESTING_TIME_END
filtered_times = filter(is_steady_state, start_end_times)
filtered_times.sort()
print 'Raw data size:', len(start_end_times),
print 'Data at steady state:', len(filtered_times)
# Figure out the actual gaps in milliseconds.
start_time_list = [start for (start, _) in filtered_times]
gap_list = []
for index in range(0, len(start_time_list) - 1):
gap = start_time_list[index + 1] - start_time_list[index]
gap_list.append(gap * 1000.0)
gap_list.sort()
print 'Client gap: (mean, stdev) =', util.get_mean_and_stdev(gap_list),
print 'median =', gap_list[len(gap_list)/2]
# Save start_time list and gap list.
with open('data/start_times.csv', 'w') as start_time_f:
for start_time_v in start_time_list:
print >> start_time_f, '%.8f' % start_time_v
with open('data/gaps.csv', 'w') as gap_f:
for (v, p) in util.make_cdf_table(gap_list):
print >> gap_f, '%f,%f' % (v, p)
# Calculate latency and bandwidth.
latency_list = []
bandwidth_list = []
for (start_time, end_time) in filtered_times:
if end_time is None:
latency = 1000
bandwidth = 0
else:
latency = end_time - start_time # seconds
bandwidth = DATA_LENGTH / latency # Bytes/s
latency_list.append(latency * 1000.0) # milliseconds
bandwidth_list.append(bandwidth * 8.0 / 1000000.0) # Mbps
# Write to file.
print 'Writing to data/async_redis_latency.csv...'
with open('data/async_redis_latency.csv', 'w') as f:
for (v, p) in util.make_cdf_table(latency_list):
print >> f, '%.10f,%.10f' % (v, p)
print 'Writing to data/async_redis_bw.csv...'
with open('data/async_redis_bw.csv', 'w') as f:
for (v, p) in util.make_cdf_table(bandwidth_list):
print >> f, '%.10f,%.10f' % (v, p)
# Analyze timings of OF events.
subprocess.call('cp of_timings.csv data/; cp /tmp/client.pcap /tmp/server.pcap data/', shell=True)
import timing_analysis
timing_analysis.main('data/client.pcap', 'data/of_timings.csv', 'data/server.pcap')
def run(packet_per_second=100, pkt_size=1500, run_time=220):
"""
Returns (pktgen_pps, pkt_in_pps, flow_mod_pps, flow_mod_pps_stdev, pkt_out_pps), where
pps_in is the actual number of packets/sec of pktgen, and flow_mod_pps and
flow_mod_pps_stdev are the mean and stdev pps of successful flow
installations at steady state.
"""
util.ping_test()
switch = Switch(config.active_config)
switch.reset_flow_table()
# Initialize the experimental controller so that POX would have the
# necessary settings.
control_client = ExpControlClient('mumu.ucsd.edu')
control_client.execute(['RESET'])
control_client.execute(['SET', 'flow_stat_interval', 20])
control_client.execute(['SET', 'install_bogus_rules', True])
control_client.execute(['SET', 'emulate_hp_switch', True])
# Start capturing packets.
tcpdump = Tcpdump(config.active_config)
tcpdump.start()
tcpdump_start_time = time.time()
# Start firing packets.
pktgen = Pktgen(config.active_config)
gap = 1.0 / packet_per_second
pkt_count = int(run_time * packet_per_second)
pktgen.low_level_start(pkt_count=pkt_count, pkt_size=pkt_size,
gap_ns=gap*1000*1000*1000, flow_count=1)
pktgen_start_time = time.time()
flow_mod_pps_list = []
# How fast were rules successfully written into the hardware table? We take
# statistics at steady state. Also display flow statistics once in a while.
last_stat_time = [0]
def callback(t_left):
flow_stat_dict = control_client.execute(['GET', 'flow_count_dict'])
for stat_time in sorted(flow_stat_dict.keys()):
if stat_time > last_stat_time[0]:
last_stat_time[0] = stat_time
flow_count = flow_stat_dict[stat_time]
print t_left, 'seconds left, with flows', flow_count
if pktgen_start_time + 60 <= time.time() <= pktgen_start_time + 180:
flow_mod_pps_list.append(flow_count / 10.0)
# Check the stat every 20 seconds.
util.callback_sleep(run_time, callback, interval=20)
# How fast were packets actually generated?
pktgen_result = pktgen.stop_and_get_result()
pktgen_pps = pktgen_result.sent_pkt_count / pktgen_result.running_time
# How fast were pkt_out events?
tcpdump_end_time = time.time()
tcpdump_result = tcpdump.stop_and_get_result()
pkt_out_pps = (tcpdump_result.dropped_pkt_count + tcpdump_result.recvd_pkt_count) / \
(tcpdump_end_time - tcpdump_start_time)
# Calculate the mean and stdev of successful flow_mod pps.
(flow_mod_pps, flow_mod_pps_stdev) = util.get_mean_and_stdev(flow_mod_pps_list)
# How fast were pkt_in events arriving?
pkt_in_count = control_client.execute(['GET', 'pkt_in_count'])
pkt_in_start_time = control_client.execute(['GET', 'pkt_in_start_time'])
pkt_in_end_time = control_client.execute(['GET', 'pkt_in_end_time'])
pkt_in_pps = pkt_in_count / (pkt_in_end_time - pkt_in_start_time)
return (pktgen_pps, pkt_in_pps, flow_mod_pps, flow_mod_pps_stdev, pkt_out_pps)