November 2011 - tor-commits - lists.torproject.org

[torflow/master] Bug #4268: Remove sleeping from bw auths
by mikeperry＠torproject.org 17 Nov '11

17 Nov '11

commit 09ac2ca82478f1f1cd2a776a936feecd5d7e9c7a Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Wed Nov 9 14:00:50 2011 -0800 Bug #4268: Remove sleeping from bw auths Code is still there, just disabling via config, and adding a check to ensure the time isn't negative. --- NetworkScanners/BwAuthority/bwauthority_child.py | 5 ++++- .../BwAuthority/data/scanner.1/bwauthority.cfg | 2 +- .../BwAuthority/data/scanner.2/bwauthority.cfg | 2 +- .../BwAuthority/data/scanner.3/bwauthority.cfg | 2 +- .../BwAuthority/data/scanner.4/bwauthority.cfg | 2 +- 5 files changed, 8 insertions(+), 5 deletions(-) diff --git a/NetworkScanners/BwAuthority/bwauthority_child.py b/NetworkScanners/BwAuthority/bwauthority_child.py index 1527a11..c42e163 100755 --- a/NetworkScanners/BwAuthority/bwauthority_child.py +++ b/NetworkScanners/BwAuthority/bwauthority_child.py @@ -216,7 +216,10 @@ def speedrace(hdlr, start_pct, stop_pct, circs_per_node, save_every, out_dir, t0 = time.time() if sleep_start <= t0 and t0 <= sleep_stop: plog("NOTICE", "It's bedtime. Sleeping for "+str(round((sleep_stop-t0)/3600.0,1))+"h") - time.sleep(sleep_stop - t0) + try: + time.sleep(sleep_stop - t0) + except: + pass t0 = time.time() hdlr.new_exit() diff --git a/NetworkScanners/BwAuthority/data/scanner.1/bwauthority.cfg b/NetworkScanners/BwAuthority/data/scanner.1/bwauthority.cfg index abd8e1e..5d986d0 100644 --- a/NetworkScanners/BwAuthority/data/scanner.1/bwauthority.cfg +++ b/NetworkScanners/BwAuthority/data/scanner.1/bwauthority.cfg @@ -25,4 +25,4 @@ circs_per_node = 5 min_streams = 1 max_fetch_time = 300 sleep_start = 01:30 -sleep_stop = 04:30 +sleep_stop = 01:30 diff --git a/NetworkScanners/BwAuthority/data/scanner.2/bwauthority.cfg b/NetworkScanners/BwAuthority/data/scanner.2/bwauthority.cfg index 9b7b591..a9544a1 100644 --- a/NetworkScanners/BwAuthority/data/scanner.2/bwauthority.cfg +++ b/NetworkScanners/BwAuthority/data/scanner.2/bwauthority.cfg @@ -25,4 +25,4 @@ circs_per_node = 5 min_streams = 1 max_fetch_time = 300 sleep_start = 01:30 -sleep_stop = 04:30 +sleep_stop = 01:30 diff --git a/NetworkScanners/BwAuthority/data/scanner.3/bwauthority.cfg b/NetworkScanners/BwAuthority/data/scanner.3/bwauthority.cfg index 51dfdc5..7999781 100644 --- a/NetworkScanners/BwAuthority/data/scanner.3/bwauthority.cfg +++ b/NetworkScanners/BwAuthority/data/scanner.3/bwauthority.cfg @@ -25,4 +25,4 @@ circs_per_node = 5 min_streams = 1 max_fetch_time = 300 sleep_start = 01:30 -sleep_stop = 04:30 +sleep_stop = 01:30 diff --git a/NetworkScanners/BwAuthority/data/scanner.4/bwauthority.cfg b/NetworkScanners/BwAuthority/data/scanner.4/bwauthority.cfg index 94342a2..c21be4f 100644 --- a/NetworkScanners/BwAuthority/data/scanner.4/bwauthority.cfg +++ b/NetworkScanners/BwAuthority/data/scanner.4/bwauthority.cfg @@ -25,4 +25,4 @@ circs_per_node = 5 min_streams = 1 max_fetch_time = 300 sleep_start = 01:30 -sleep_stop = 04:30 +sleep_stop = 01:30

1 0

[torflow/master] Differentiate measured and update timestamps.
by mikeperry＠torproject.org 17 Nov '11

17 Nov '11

commit e87c3e379992af11d23ed7b916ad65b19e6eda56 Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Sun Nov 13 18:25:36 2011 -0800 Differentiate measured and update timestamps. This helps us to track Guard node feedback/non-feedback better. We may also end up using this to determine T_d and possibly T_i. Also improve some log messages. --- NetworkScanners/BwAuthority/aggregate.py | 59 ++++++++++++++++++++---------- 1 files changed, 40 insertions(+), 19 deletions(-) diff --git a/NetworkScanners/BwAuthority/aggregate.py b/NetworkScanners/BwAuthority/aggregate.py index 03822de..ca202bb 100755 --- a/NetworkScanners/BwAuthority/aggregate.py +++ b/NetworkScanners/BwAuthority/aggregate.py @@ -47,9 +47,11 @@ NODE_CAP = 0.05 MIN_REPORT = 60 # Percent of the network we must measure before reporting # Keep most measurements in consideration. The code below chooses -# the most recent one. 15 days is just to stop us from choking up +# the most recent one. 28 days is just to stop us from choking up # all the CPU once these things run for a year or so. -MAX_AGE = 60*60*24*15 +# Note that the Guard measurement interval of 2 weeks means that this +# value can't get much below that. +MAX_AGE = 2*GUARD_SAMPLE_RATE # If the resultant scan file is older than 1.5 days, something is wrong MAX_SCAN_AGE = 60*60*24*1.5 @@ -81,7 +83,6 @@ class Node: self.pid_bw = 0 self.pid_error = 0 self.prev_error = 0 - self.prev_measured_at = 0 self.pid_error_sum = 0 self.pid_delta = 0 self.ratio = None @@ -96,6 +97,7 @@ class Node: self.desc_bw = 0 self.circ_fail_rate = 0 self.strm_fail_rate = 0 + self.updated_at = 0 def revert_to_vote(self, vote): self.new_bw = vote.bw*1000 @@ -108,7 +110,6 @@ class Node: # Derivative of error for pid control def get_pid_bw(self, prev_vote, kp, ki, kd, kidecay): self.prev_error = prev_vote.pid_error - self.prev_measured_at = prev_vote.measured_at # We decay the interval each round to keep it bounded. # This decay is non-standard. We do it to avoid overflow self.pid_error_sum = prev_vote.pid_error_sum*kidecay + self.pid_error @@ -127,7 +128,7 @@ class Node: # Rate of change in error from the last measurement sample def d_error_dt(self): - if self.prev_measured_at == 0 or self.prev_error == 0: + if self.prev_error == 0: self.pid_delta = 0 else: self.pid_delta = self.pid_error - self.prev_error @@ -139,7 +140,7 @@ class Node: self.idhex = line.idhex self.nick = line.nick if line.measured_at > self.measured_at: - self.measured_at = line.measured_at + self.measured_at = self.updated_at = line.measured_at self.strm_bw = line.strm_bw self.filt_bw = line.filt_bw self.ns_bw = line.ns_bw @@ -185,6 +186,12 @@ class Vote: self.pid_error = 0 self.pid_delta = 0 self.pid_error_sum = 0 + try: + self.updated_at = int(re.search("[\s]*updated_at=([\S]+)[\s]*", line).group(1)) + except: + plog("INFO", "No updated_at field for "+self.nick+"="+self.idhex) + self.updated_at = self.measured_at + class VoteSet: def __init__(self, filename): @@ -434,19 +441,23 @@ def main(argv): if n.idhex in prev_votes.vote_map and n.idhex in prev_consensus: if "Guard" in prev_consensus[n.idhex].flags and \ "Exit" not in prev_consensus[n.idhex].flags: - guard_cnt += 1 - guard_measure_time += (n.measured_at - \ - prev_votes.vote_map[n.idhex].measured_at) + if n.measured_at != prev_votes.vote_map[n.idhex].measured_at: + guard_cnt += 1 + guard_measure_time += (n.measured_at - \ + prev_votes.vote_map[n.idhex].measured_at) else: - node_cnt += 1 - node_measure_time += (n.measured_at - \ - prev_votes.vote_map[n.idhex].measured_at) + if n.updated_at != prev_votes.vote_map[n.idhex].updated_at: + node_cnt += 1 + node_measure_time += (n.updated_at - \ + prev_votes.vote_map[n.idhex].updated_at) + # TODO: We may want to try to use this info to autocompute T_d and + # maybe T_i? if node_cnt > 0: - plog("INFO", "Average node measurement interval: "+str(node_measure_time/node_cnt)) + plog("INFO", "Avg of "+str(node_cnt)+" node update intervals: "+str((node_measure_time/node_cnt)/3600.0)) if guard_cnt > 0: - plog("INFO", "Average gaurd measurement interval: "+str(guard_measure_time/guard_cnt)) + plog("INFO", "Avg of "+str(guard_cnt)+" guard measurement interval: "+str((guard_measure_time/guard_cnt)/3600.0)) tot_net_bw = 0 for n in nodes.itervalues(): @@ -484,6 +495,7 @@ def main(argv): # new measurement against the previous vote. n.new_bw = prev_votes.vote_map[n.idhex].pid_bw + \ cs_junk.K_p*prev_votes.vote_map[n.idhex].pid_bw*pid_error + else: # Everyone else should be pretty instantenous to respond. # Full feedback should be fine for them (we hope), @@ -510,7 +522,7 @@ def main(argv): # Reset values. Don't vote/sample this measurement round. n.revert_to_vote(prev_votes.vote_map[n.idhex]) else: # No prev vote, pure consensus feedback this round - n.new_bw = n.ns_bw + K_p*n.ns_bw*n.pid_error + n.new_bw = n.ns_bw + cs_junk.K_p*n.ns_bw*n.pid_error n.pid_error_sum = n.pid_error n.pid_bw = n.new_bw plog("INFO", "No prev vote for node "+n.nick+": Consensus feedback") @@ -546,6 +558,10 @@ def main(argv): if n.new_bw >= 0xffffffff*1000: plog("WARN", "Bandwidth of node "+n.nick+"="+n.idhex+" exceeded maxint32: "+str(n.new_bw)) n.new_bw = 0xffffffff*1000 + if cs_junk.T_i > 0 and math.fabs(n.pid_error_sum) > \ + math.fabs(2*cs_junk.T_i*n.pid_error/cs_junk.T_i_decay): + plog("NOTICE", "Large pid_error_sum for node "+n.idhex+"="+n.nick+": "+ + str(n.pid_error_sum)+" vs "+str(n.pid_error)) if n.new_bw > tot_net_bw*NODE_CAP: plog("INFO", "Clipping extremely fast node "+n.idhex+"="+n.nick+ " at "+str(100*NODE_CAP)+"% of network capacity ("+ @@ -561,11 +577,15 @@ def main(argv): plog("INFO", "New node "+n.idhex+"="+n.nick+" has bandwidth < 0: "+str(n.new_bw)) n.new_bw = 1 - # WTF is going on here? - oldest_timestamp = min(map(lambda n: n.measured_at, + oldest_measured = min(map(lambda n: n.measured_at, + filter(lambda n: n.idhex in prev_consensus, + nodes.itervalues()))) + plog("INFO", "Oldest measured node: "+time.ctime(oldest_measured)) + + oldest_updated = min(map(lambda n: n.updated_at, filter(lambda n: n.idhex in prev_consensus, nodes.itervalues()))) - plog("INFO", "Oldest measured node: "+time.ctime(oldest_timestamp)) + plog("INFO", "Oldest updated node: "+time.ctime(oldest_updated)) missed_nodes = 0.0 for n in prev_consensus.itervalues(): @@ -604,7 +624,8 @@ def main(argv): # FIXME: Split out debugging data for n in n_print: if not n.ignore: - out.write("node_id="+n.idhex+" bw="+str(base10_round(n.new_bw))+" nick="+n.nick+ " measured_at="+str(int(n.measured_at))+" pid_error="+str(n.pid_error)+" pid_error_sum="+str(n.pid_error_sum)+" pid_bw="+str(int(n.pid_bw))+" pid_delta="+str(n.pid_delta)+" circ_fail="+str(n.circ_fail_rate)+"\n") + # Turns out str() is more accurate than %lf + out.write("node_id="+n.idhex+" bw="+str(base10_round(n.new_bw))+" nick="+n.nick+ " measured_at="+str(int(n.measured_at))+" updated_at="+str(int(n.updated_at))+" pid_error="+str(n.pid_error)+" pid_error_sum="+str(n.pid_error_sum)+" pid_bw="+str(int(n.pid_bw))+" pid_delta="+str(n.pid_delta)+" circ_fail="+str(n.circ_fail_rate)+"\n") out.close() write_file_list(argv[1])

1 0

[torflow/master] Move more parameters to consensus.
by mikeperry＠torproject.org 17 Nov '11

17 Nov '11

commit 13aa3b4268a98a07e76e8ba9cb8cfccac39ff2c0 Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Sun Nov 13 17:45:01 2011 -0800 Move more parameters to consensus. --- NetworkScanners/BwAuthority/aggregate.py | 103 +++++++++++++++++++++++------- 1 files changed, 79 insertions(+), 24 deletions(-) diff --git a/NetworkScanners/BwAuthority/aggregate.py b/NetworkScanners/BwAuthority/aggregate.py index 1746256..9bed1d0 100755 --- a/NetworkScanners/BwAuthority/aggregate.py +++ b/NetworkScanners/BwAuthority/aggregate.py @@ -23,22 +23,25 @@ IGNORE_GUARDS = 0 # rate for guard nodes GUARD_SAMPLE_RATE = 2*7*24*60*60 # 2wks -# PID constants -# See https://en.wikipedia.org/wiki/PID_controller#Ideal_versus_standard_PID_form +# PID constant defaults. May be overridden by consensus +# https://en.wikipedia.org/wiki/PID_controller#Ideal_versus_standard_PID_form K_p = 1.0 # We expect to correct steady state error in 5 samples (guess) T_i = 5.0 +# T_i_decay is a weight factor to govern how fast integral sums +# decay. For the values of T_i that we care about, T_i_decay represents +# the fraction of integral sum that is eliminated after T_i sample rounds. +# This decay is non-standard, but we do it to avoid overflow +T_i_decay = 0.5 + # We can only expect to predict less than one sample into the future, as # after 1 sample, clients will have migrated -# FIXME: Our prediction ability is a function of the consensus time +# FIXME: Our prediction ability is a function of the consensus uptake time # vs measurement rate T_d = 0.5 -K_i = K_p/T_i -K_d = K_p*T_d - NODE_CAP = 0.05 MIN_REPORT = 60 # Percent of the network we must measure before reporting @@ -94,6 +97,7 @@ class Node: self.circ_fail_rate = 0 self.strm_fail_rate = 0 + # FIXME: Need to set pid_error_sum.. It is getting lost when we don't vote def revert_to_vote(self, vote): self.new_bw = vote.bw*1000 self.pid_bw = vote.pid_bw @@ -101,17 +105,17 @@ class Node: self.measured_at = vote.measured_at # Derivative of error for pid control - def get_pid_bw(self, prev_vote, kp): + def get_pid_bw(self, prev_vote, kp, ki, kd, kidecay): self.prev_error = prev_vote.pid_error self.prev_measured_at = prev_vote.measured_at - # We decay the interval by 1/T_i each round to keep it bounded. - # This is non-standard - self.pid_error_sum = prev_vote.pid_error_sum*(1 - 1.0/T_i) + self.pid_error + # We decay the interval each round to keep it bounded. + # This decay is non-standard. We do it to avoid overflow + self.pid_error_sum = prev_vote.pid_error_sum*kidecay + self.pid_error self.pid_bw = self.ns_bw \ - + kp*(self.ns_bw*self.pid_error \ - + self.ns_bw*self.integral_error()/T_i \ - + self.ns_bw*self.d_error_dt()*T_d) + + kp*self.ns_bw*self.pid_error \ + + ki*self.ns_bw*self.integral_error() \ + + kd*self.ns_bw*self.d_error_dt() return self.pid_bw # Time-weighted sum of error per unit of time (measurement sample) @@ -139,8 +143,7 @@ class Node: self.filt_bw = line.filt_bw self.ns_bw = line.ns_bw self.desc_bw = line.desc_bw - # XXX: Temporary test - self.circ_fail_rate = 0.0 #line.circ_fail_rate + self.circ_fail_rate = line.circ_fail_rate self.strm_fail_rate = line.strm_fail_rate class Line: @@ -197,16 +200,54 @@ class ConsensusJunk: def __init__(self, c): cs_bytes = c.sendAndRecv("GETINFO dir/status-vote/current/consensus\r\n")[0][2] self.bwauth_pid_control = False + self.use_circ_fails = False + self.use_best_ratio = False + + self.K_p = K_p + self.T_i = T_i + self.T_d = T_d + self.T_i_decay = T_i_decay + try: cs_params = re.search("^params ((?:[\S]+=[\d]+[\s]?)+)", cs_bytes, re.M).group(1).split() for p in cs_params: if p == "bwauthpid=1": self.bwauth_pid_control = True + elif p == "bwauthcircs=1": + self.use_circ_fails = True + plog("INFO", "Counting circuit failures") + elif p == "bwauthbestratio=1": + self.use_best_ratio = True + plog("INFO", "Choosing larger of sbw vs fbw") + elif p.startswith("bwauthkp="): + self.K_p = int(p.split("=")[1])/10000.0 + plog("INFO", "Got K_p=%f from consensus." % self.K_p) + elif p.startswith("bwauthti="): + self.T_i = (int(p.split("=")[1])/10000.0) + plog("INFO", "Got T_i=%f from consensus." % self.T_i) + elif p.startswith("bwauthtd="): + self.T_d = (int(p.split("=")[1])/10000.0) + plog("INFO", "Got T_d=%f from consensus." % self.T_d) + elif p.startswith("bwauthtidecay="): + self.T_i_decay = (int(p.split("=")[1])/10000.0) + plog("INFO", "Got T_i_decay=%f from consensus." % self.T_i_decay) except: plog("NOTICE", "Bw auth PID control disabled due to parse error.") traceback.print_exc() + if self.T_i == 0: + self.K_i = 0 + self.K_i_decay = 0 + else: + self.K_i = self.K_p/self.T_i + self.K_i_decay = (1.0-self.T_i_decay/self.T_i) + + self.K_d = self.K_p*self.T_d + + plog("INFO", "Got K_p=%f K_i=%f K_d=%f K_i_decay=%f" % + (self.K_p, self.K_i, self.K_d, self.K_i_decay)) + self.bw_weights = {} try: bw_weights = re.search("^bandwidth-weights ((?:[\S]+=[\d]+[\s]?)+)", @@ -355,6 +396,11 @@ def main(argv): plog("NOTICE", "No scan results yet.") sys.exit(1) + if not cs_junk.use_circ_fails: + plog("INFO", "Ignoring circuit failures") + for n in nodes.itervalues(): + n.circ_fail_rate = 0.0 + if cs_junk.bwauth_pid_control: # Penalize nodes for circuit failure: it indicates CPU pressure # TODO: Potentially penalize for stream failure, if we run into @@ -406,10 +452,7 @@ def main(argv): if cs_junk.bwauth_pid_control: # Penalize nodes for circ failure rate - # n.pid_error = (n.filt_bw*(1.0-n.circ_fail_rate) - true_filt_avg)/true_filt_avg - - # FIXME: For now, let's try this larger-ratio thing.. - if n.sbw_ratio > n.fbw_ratio: + if cs_junk.use_best_ratio and n.sbw_ratio > n.fbw_ratio: n.pid_error = (n.strm_bw*(1.0-n.circ_fail_rate) - true_strm_avg)/true_strm_avg else: n.pid_error = (n.filt_bw*(1.0-n.circ_fail_rate) - true_filt_avg)/true_filt_avg @@ -426,9 +469,14 @@ def main(argv): # Do full feedback if our previous vote > 2.5 weeks old if n.idhex not in prev_votes.vote_map or \ n.measured_at - prev_votes.vote_map[n.idhex].measured_at > GUARD_SAMPLE_RATE: - n.new_bw = n.get_pid_bw(prev_votes.vote_map[n.idhex], K_p) + n.new_bw = n.get_pid_bw(prev_votes.vote_map[n.idhex], + cs_junk.K_p, + cs_junk.K_i, + cs_junk.K_d, + cs_junk.K_i_decay) else: pid_error = n.pid_error + # FIXME: We possibly lose the pid_error_sum here n.revert_to_vote(prev_votes.vote_map[n.idhex]) # Don't use feedback here, but we might as well use our # new measurement against the previous vote. @@ -436,11 +484,11 @@ def main(argv): # This should no longer happen plog("NOTICE", "Zero bw for Guard node "+n.nick+"="+n.idhex) n.new_bw = prev_votes.vote_map[n.idhex].bw + \ - K_p*prev_votes.vote_map[n.idhex].bw*pid_error + cs_junk.K_p*prev_votes.vote_map[n.idhex].bw*pid_error n.pid_bw = n.new_bw else: n.new_bw = prev_votes.vote_map[n.idhex].pid_bw + \ - K_p*prev_votes.vote_map[n.idhex].pid_bw*pid_error + cs_junk.K_p*prev_votes.vote_map[n.idhex].pid_bw*pid_error else: # Everyone else should be pretty instantenous to respond. # Full feedback should be fine for them (we hope), @@ -453,9 +501,16 @@ def main(argv): ("Guard" in prev_consensus[n.idhex].flags \ and "Exit" in prev_consensus[n.idhex].flags): n.new_bw = n.get_pid_bw(prev_votes.vote_map[n.idhex], - K_p*(1.0-cs_junk.bw_weights["Wgd"])) + cs_junk.K_p*(1.0-cs_junk.bw_weights["Wgd"]), + cs_junk.K_i, + cs_junk.K_d, + cs_junk.K_i_decay) else: - n.new_bw = n.get_pid_bw(prev_votes.vote_map[n.idhex], K_p) + n.new_bw = n.get_pid_bw(prev_votes.vote_map[n.idhex], + cs_junk.K_p, + cs_junk.K_i, + cs_junk.K_d, + cs_junk.K_i_decay) else: # Reset values. Don't vote/sample this measurement round. n.revert_to_vote(prev_votes.vote_map[n.idhex])

1 0

[torflow/master] Prevent file size list from double-listing percentiles.
by mikeperry＠torproject.org 17 Nov '11

17 Nov '11

commit a47e441e0a0e565e2101c1eb738aab6a800c132e Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Mon Nov 14 17:10:45 2011 -0800 Prevent file size list from double-listing percentiles. If two different file sizes serve one percentile, go with the smaller file size. --- NetworkScanners/BwAuthority/aggregate.py | 9 ++++++++- 1 files changed, 8 insertions(+), 1 deletions(-) diff --git a/NetworkScanners/BwAuthority/aggregate.py b/NetworkScanners/BwAuthority/aggregate.py index ca202bb..fb5cc7a 100755 --- a/NetworkScanners/BwAuthority/aggregate.py +++ b/NetworkScanners/BwAuthority/aggregate.py @@ -280,17 +280,24 @@ def write_file_list(datadir): file_sizes.sort(reverse=True) node_fbws.sort() prev_size = file_sizes[-1] + prev_pct = 0 i = 0 # The idea here is to grab the largest file size such # that 5*bw < file, and do this for each file size. for bw in node_fbws: i += 1 + pct = 100-(100*i)/len(node_fbws) + # If two different file sizes serve one percentile, go with the + # smaller file size (ie skip this one) + if pct == prev_pct: + continue for f in xrange(len(file_sizes)): if bw > file_sizes[f]*1024 and file_sizes[f] > prev_size: next_f = max(f-1,0) - file_pairs.append((100-(100*i)/len(node_fbws),files[file_sizes[next_f]])) + file_pairs.append((pct,files[file_sizes[next_f]])) prev_size = file_sizes[f] + prev_pct = pct break file_pairs.reverse()

1 0

[torflow/master] Retain pid_error_sum and pid_delta between votes.
by mikeperry＠torproject.org 17 Nov '11

17 Nov '11

commit cd75fc9a01e99d4df34329e6846a640a1a401078 Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Sun Nov 13 18:07:41 2011 -0800 Retain pid_error_sum and pid_delta between votes. pid_delta is stored for informational purposes only. However, missing pid_error_sum for Guard measurements w/o feedback was a bug. --- NetworkScanners/BwAuthority/aggregate.py | 31 +++++++++++++---------------- 1 files changed, 14 insertions(+), 17 deletions(-) diff --git a/NetworkScanners/BwAuthority/aggregate.py b/NetworkScanners/BwAuthority/aggregate.py index 9bed1d0..03822de 100755 --- a/NetworkScanners/BwAuthority/aggregate.py +++ b/NetworkScanners/BwAuthority/aggregate.py @@ -83,7 +83,7 @@ class Node: self.prev_error = 0 self.prev_measured_at = 0 self.pid_error_sum = 0 - self.derror_dt = 0 + self.pid_delta = 0 self.ratio = None self.new_bw = None self.change = None @@ -97,11 +97,12 @@ class Node: self.circ_fail_rate = 0 self.strm_fail_rate = 0 - # FIXME: Need to set pid_error_sum.. It is getting lost when we don't vote def revert_to_vote(self, vote): self.new_bw = vote.bw*1000 self.pid_bw = vote.pid_bw self.pid_error = vote.pid_error + self.pid_error_sum = vote.pid_error_sum + self.pid_delta = vote.pid_delta self.measured_at = vote.measured_at # Derivative of error for pid control @@ -127,10 +128,10 @@ class Node: # Rate of change in error from the last measurement sample def d_error_dt(self): if self.prev_measured_at == 0 or self.prev_error == 0: - self.derror_dt = 0 + self.pid_delta = 0 else: - self.derror_dt = self.pid_error - self.prev_error - return self.derror_dt + self.pid_delta = self.pid_error - self.prev_error + return self.pid_delta def add_line(self, line): if self.idhex and self.idhex != line.idhex: @@ -176,11 +177,13 @@ class Vote: try: self.pid_error = float(re.search("[\s]*pid_error=([\S]+)[\s]*", line).group(1)) self.pid_error_sum = float(re.search("[\s]*pid_error_sum=([\S]+)[\s]*", line).group(1)) + self.pid_delta = float(re.search("[\s]*pid_delta=([\S]+)[\s]*", line).group(1)) self.pid_bw = float(re.search("[\s]*pid_bw=([\S]+)[\s]*", line).group(1)) except: plog("NOTICE", "No previous PID data.") self.pid_bw = self.bw self.pid_error = 0 + self.pid_delta = 0 self.pid_error_sum = 0 class VoteSet: @@ -476,18 +479,10 @@ def main(argv): cs_junk.K_i_decay) else: pid_error = n.pid_error - # FIXME: We possibly lose the pid_error_sum here n.revert_to_vote(prev_votes.vote_map[n.idhex]) # Don't use feedback here, but we might as well use our # new measurement against the previous vote. - if prev_votes.vote_map[n.idhex].pid_bw == 0: - # This should no longer happen - plog("NOTICE", "Zero bw for Guard node "+n.nick+"="+n.idhex) - n.new_bw = prev_votes.vote_map[n.idhex].bw + \ - cs_junk.K_p*prev_votes.vote_map[n.idhex].bw*pid_error - n.pid_bw = n.new_bw - else: - n.new_bw = prev_votes.vote_map[n.idhex].pid_bw + \ + n.new_bw = prev_votes.vote_map[n.idhex].pid_bw + \ cs_junk.K_p*prev_votes.vote_map[n.idhex].pid_bw*pid_error else: # Everyone else should be pretty instantenous to respond. @@ -553,8 +548,10 @@ def main(argv): n.new_bw = 0xffffffff*1000 if n.new_bw > tot_net_bw*NODE_CAP: plog("INFO", "Clipping extremely fast node "+n.idhex+"="+n.nick+ - " at "+str(100*NODE_CAP)+"% of network capacity (" - +str(n.new_bw)+"->"+str(int(tot_net_bw*NODE_CAP))+")") + " at "+str(100*NODE_CAP)+"% of network capacity ("+ + str(n.new_bw)+"->"+str(int(tot_net_bw*NODE_CAP))+") "+ + " pid_error="+str(n.pid_error)+ + " pid_error_sum="+str(n.pid_error_sum)) n.new_bw = int(tot_net_bw*NODE_CAP) n.pid_error_sum = 0 # Don't let unused error accumulate... if n.new_bw <= 0: @@ -607,7 +604,7 @@ def main(argv): # FIXME: Split out debugging data for n in n_print: if not n.ignore: - out.write("node_id="+n.idhex+" bw="+str(base10_round(n.new_bw))+" nick="+n.nick+ " measured_at="+str(int(n.measured_at))+" pid_error="+str(n.pid_error)+" pid_error_sum="+str(n.pid_error_sum)+" pid_bw="+str(int(n.pid_bw))+" pid_delta="+str(n.derror_dt)+" circ_fail="+str(n.circ_fail_rate)+"\n") + out.write("node_id="+n.idhex+" bw="+str(base10_round(n.new_bw))+" nick="+n.nick+ " measured_at="+str(int(n.measured_at))+" pid_error="+str(n.pid_error)+" pid_error_sum="+str(n.pid_error_sum)+" pid_bw="+str(int(n.pid_bw))+" pid_delta="+str(n.pid_delta)+" circ_fail="+str(n.circ_fail_rate)+"\n") out.close() write_file_list(argv[1])

1 0

[torflow/master] Update pythonpath to modern versions.
by mikeperry＠torproject.org 17 Nov '11

17 Nov '11

commit f261978b21b987b3304eb3d486a82949f7d46def Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Mon Nov 14 17:35:33 2011 -0800 Update pythonpath to modern versions. --- NetworkScanners/BwAuthority/run_scan.sh | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-) diff --git a/NetworkScanners/BwAuthority/run_scan.sh b/NetworkScanners/BwAuthority/run_scan.sh index b3e1355..bc25489 100755 --- a/NetworkScanners/BwAuthority/run_scan.sh +++ b/NetworkScanners/BwAuthority/run_scan.sh @@ -6,7 +6,7 @@ SCANNER_COUNT=4 # This tor must have the w status line fix as well as the stream bw fix # Ie git master or 0.2.2.x TOR_EXE=../../../tor.git/src/or/tor -PYTHONPATH=../../../SQLAlchemy-0.5.5/lib:../../../Elixir-0.6.1/ +PYTHONPATH=../../../SQLAlchemy-0.7.3/lib:../../../Elixir-0.7.1/ for n in `seq $SCANNER_COUNT`; do PIDFILE=./data/scanner.${n}/bwauthority.pid

1 0

[torflow/master] Update scanner cutoffs and timeouts in config.
by mikeperry＠torproject.org 17 Nov '11

17 Nov '11

commit 781e4d11eb38deaf9b509438b7b7a1a771e5efa1 Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Mon Nov 14 18:10:54 2011 -0800 Update scanner cutoffs and timeouts in config. --- .../BwAuthority/data/scanner.1/bwauthority.cfg | 4 ++-- .../BwAuthority/data/scanner.2/bwauthority.cfg | 4 ++-- .../BwAuthority/data/scanner.3/bwauthority.cfg | 2 +- .../BwAuthority/data/scanner.4/bwauthority.cfg | 2 +- 4 files changed, 6 insertions(+), 6 deletions(-) diff --git a/NetworkScanners/BwAuthority/data/scanner.1/bwauthority.cfg b/NetworkScanners/BwAuthority/data/scanner.1/bwauthority.cfg index 5d986d0..c15f5c5 100644 --- a/NetworkScanners/BwAuthority/data/scanner.1/bwauthority.cfg +++ b/NetworkScanners/BwAuthority/data/scanner.1/bwauthority.cfg @@ -18,11 +18,11 @@ db_url = #db_url = postgresql://bwscanner:password@127.0.0.1/BwScan1 tor_dir = ./data/tor start_pct = 0 -stop_pct = 12 +stop_pct = 10 save_every = 0 nodes_per_slice = 50 circs_per_node = 5 min_streams = 1 -max_fetch_time = 300 +max_fetch_time = 600 sleep_start = 01:30 sleep_stop = 01:30 diff --git a/NetworkScanners/BwAuthority/data/scanner.2/bwauthority.cfg b/NetworkScanners/BwAuthority/data/scanner.2/bwauthority.cfg index a9544a1..68d5f9c 100644 --- a/NetworkScanners/BwAuthority/data/scanner.2/bwauthority.cfg +++ b/NetworkScanners/BwAuthority/data/scanner.2/bwauthority.cfg @@ -17,8 +17,8 @@ db_url = #db_url = mysql+mysqldb://bwscanner:password@127.0.0.1/BwScan2 #db_url = postgresql://bwscanner:password@127.0.0.1/BwScan2 tor_dir = ./data/tor -start_pct = 12 -stop_pct = 35 +start_pct = 10 +stop_pct = 40 save_every = 0 nodes_per_slice = 50 circs_per_node = 5 diff --git a/NetworkScanners/BwAuthority/data/scanner.3/bwauthority.cfg b/NetworkScanners/BwAuthority/data/scanner.3/bwauthority.cfg index 7999781..7ee1808 100644 --- a/NetworkScanners/BwAuthority/data/scanner.3/bwauthority.cfg +++ b/NetworkScanners/BwAuthority/data/scanner.3/bwauthority.cfg @@ -17,7 +17,7 @@ db_url = #db_url = mysql+mysqldb://bwscanner:password@127.0.0.1/BwScan3 #db_url = postgresql://bwscanner:password@127.0.0.1/BwScan3 tor_dir = ./data/tor -start_pct =35 +start_pct =30 stop_pct = 60 save_every = 0 nodes_per_slice = 50 diff --git a/NetworkScanners/BwAuthority/data/scanner.4/bwauthority.cfg b/NetworkScanners/BwAuthority/data/scanner.4/bwauthority.cfg index c21be4f..d0db161 100644 --- a/NetworkScanners/BwAuthority/data/scanner.4/bwauthority.cfg +++ b/NetworkScanners/BwAuthority/data/scanner.4/bwauthority.cfg @@ -17,7 +17,7 @@ db_url = #db_url = mysql+mysqldb://bwscanner:password@127.0.0.1/BwScan4 #db_url = postgresql://bwscanner:password@127.0.0.1/BwScan4 tor_dir = ./data/tor -start_pct =60 +start_pct =70 stop_pct = 100 save_every = 0 nodes_per_slice = 50

1 0

[torflow/master] Add some comments about potential timeout issues.
by mikeperry＠torproject.org 17 Nov '11

17 Nov '11

commit 1efe897f89b6ff113862dd0d418a17873962b836 Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Mon Nov 14 18:11:35 2011 -0800 Add some comments about potential timeout issues. --- NetworkScanners/BwAuthority/bwauthority_child.py | 7 ++++++- 1 files changed, 6 insertions(+), 1 deletions(-) diff --git a/NetworkScanners/BwAuthority/bwauthority_child.py b/NetworkScanners/BwAuthority/bwauthority_child.py index c42e163..65063a9 100755 --- a/NetworkScanners/BwAuthority/bwauthority_child.py +++ b/NetworkScanners/BwAuthority/bwauthority_child.py @@ -225,8 +225,11 @@ def speedrace(hdlr, start_pct, stop_pct, circs_per_node, save_every, out_dir, hdlr.new_exit() attempt += 1 - # FIXME: This noise is due to a difficult to find Tor bug that + # TODO: This noise is due to a difficult to find Tor bug that # causes some exits to hang forever on streams :( + # FIXME: Hrmm, should we change the reason on this? Right now, + # 7 == TIMEOUT, which means we do not count the bandwidth of this + # stream.. however, we count it as 'successful' below timer = threading.Timer(max_fetch_time, lambda: hdlr.close_streams(7)) timer.start() url = choose_url(start_pct) @@ -240,6 +243,8 @@ def speedrace(hdlr, start_pct, stop_pct, circs_per_node, save_every, out_dir, plog('WARN', 'Timer exceeded limit: ' + str(delta_build) + '\n') build_exit = hdlr.get_exit_node() + # FIXME: Timeouts get counted as 'sucessful' here, but do not + # count in the SQL stats! if ret == 1 and build_exit: successful += 1 plog('DEBUG', str(start_pct) + '-' + str(stop_pct) + '% circuit build+fetch took ' + str(delta_build) + ' for ' + str(build_exit))

1 0

[torflow/master] Submodule.
by mikeperry＠torproject.org 17 Nov '11

17 Nov '11

commit 8637ba1e6afb792c332b0a14f535f7c6d7d54c28 Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Tue Nov 15 18:45:16 2011 -0800 Submodule. --- TorCtl | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-) diff --git a/TorCtl b/TorCtl index 391b439..3de9176 160000 --- a/TorCtl +++ b/TorCtl @@ -1 +1 @@ -Subproject commit 391b4399adda91e8b68ffd6e1fa294efe846debd +Subproject commit 3de91767e208d1642a8cf89dc1b645f637abaf8e

1 0

[torflow/master] Allow desc bw to be used in feedback.
by mikeperry＠torproject.org 17 Nov '11

17 Nov '11

commit 0e9738d8d5670472ae0e503629f1bbfca3ff6d9b Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Tue Nov 15 21:06:34 2011 -0800 Allow desc bw to be used in feedback. It may be equivalent to NS bw, since the NS bw functions as an accumulator.. We may want to set T_i_decay to 0 for this code. --- NetworkScanners/BwAuthority/aggregate.py | 76 +++++++++++++++++++++-------- 1 files changed, 55 insertions(+), 21 deletions(-) diff --git a/NetworkScanners/BwAuthority/aggregate.py b/NetworkScanners/BwAuthority/aggregate.py index fb5cc7a..c55b9a1 100755 --- a/NetworkScanners/BwAuthority/aggregate.py +++ b/NetworkScanners/BwAuthority/aggregate.py @@ -88,6 +88,7 @@ class Node: self.ratio = None self.new_bw = None self.change = None + self.use_bw = -1 # measurement vars from bwauth lines self.measured_at = 0 @@ -100,24 +101,32 @@ class Node: self.updated_at = 0 def revert_to_vote(self, vote): - self.new_bw = vote.bw*1000 - self.pid_bw = vote.pid_bw - self.pid_error = vote.pid_error - self.pid_error_sum = vote.pid_error_sum - self.pid_delta = vote.pid_delta - self.measured_at = vote.measured_at - - # Derivative of error for pid control - def get_pid_bw(self, prev_vote, kp, ki, kd, kidecay): + self.copy_vote(vote) + self.pid_error = vote.pid_error # Set + self.measured_at = vote.measured_at # Set + + def copy_vote(self, vote): + self.new_bw = vote.bw*1000 # Not set yet + self.pid_bw = vote.pid_bw # Not set yet + self.pid_error_sum = vote.pid_error_sum # Not set yet + self.pid_delta = vote.pid_delta # Not set yet + + def get_pid_bw(self, prev_vote, kp, ki, kd, kidecay, update=True): + if not update: + return self.use_bw \ + + kp*self.use_bw*self.pid_error \ + + ki*self.use_bw*self.pid_error_sum \ + + kd*self.use_bw*self.pid_delta + self.prev_error = prev_vote.pid_error # We decay the interval each round to keep it bounded. # This decay is non-standard. We do it to avoid overflow self.pid_error_sum = prev_vote.pid_error_sum*kidecay + self.pid_error - self.pid_bw = self.ns_bw \ - + kp*self.ns_bw*self.pid_error \ - + ki*self.ns_bw*self.integral_error() \ - + kd*self.ns_bw*self.d_error_dt() + self.pid_bw = self.use_bw \ + + kp*self.use_bw*self.pid_error \ + + ki*self.use_bw*self.integral_error() \ + + kd*self.use_bw*self.d_error_dt() return self.pid_bw # Time-weighted sum of error per unit of time (measurement sample) @@ -212,6 +221,7 @@ class ConsensusJunk: self.bwauth_pid_control = False self.use_circ_fails = False self.use_best_ratio = False + self.use_desc_bw = False self.K_p = K_p self.T_i = T_i @@ -224,6 +234,9 @@ class ConsensusJunk: for p in cs_params: if p == "bwauthpid=1": self.bwauth_pid_control = True + elif p == "bwauthdescbw=1": + self.use_desc_bw = True + plog("INFO", "Using descriptor bandwidth") elif p == "bwauthcircs=1": self.use_circ_fails = True plog("INFO", "Counting circuit failures") @@ -472,6 +485,11 @@ def main(argv): n.sbw_ratio = n.strm_bw/true_strm_avg if cs_junk.bwauth_pid_control: + if cs_junk.use_desc_bw: + n.use_bw = n.desc_bw + else: + n.use_bw = n.ns_bw + # Penalize nodes for circ failure rate if cs_junk.use_best_ratio and n.sbw_ratio > n.fbw_ratio: n.pid_error = (n.strm_bw*(1.0-n.circ_fail_rate) - true_strm_avg)/true_strm_avg @@ -496,13 +514,28 @@ def main(argv): cs_junk.K_d, cs_junk.K_i_decay) else: - pid_error = n.pid_error - n.revert_to_vote(prev_votes.vote_map[n.idhex]) # Don't use feedback here, but we might as well use our # new measurement against the previous vote. - n.new_bw = prev_votes.vote_map[n.idhex].pid_bw + \ - cs_junk.K_p*prev_votes.vote_map[n.idhex].pid_bw*pid_error - + n.copy_vote(prev_votes.vote_map[n.idhex]) + + if cs_junk.use_desc_bw: + n.new_bw = n.get_pid_bw(prev_votes.vote_map[n.idhex], + cs_junk.K_p, + cs_junk.K_i, + cs_junk.K_d, + 0.0, False) + else: + # Use previous vote's feedback bw + n.use_bw = prev_votes.vote_map[n.idhex].pid_bw + n.new_bw = n.get_pid_bw(prev_votes.vote_map[n.idhex], + cs_junk.K_p, + 0.0, + 0.0, + 0.0, False) + + # Reset the remaining vote data.. + n.measured_at = prev_votes.vote_map[n.idhex].measured_at + n.pid_error = prev_votes.vote_map[n.idhex].pid_error else: # Everyone else should be pretty instantenous to respond. # Full feedback should be fine for them (we hope), @@ -529,7 +562,7 @@ def main(argv): # Reset values. Don't vote/sample this measurement round. n.revert_to_vote(prev_votes.vote_map[n.idhex]) else: # No prev vote, pure consensus feedback this round - n.new_bw = n.ns_bw + cs_junk.K_p*n.ns_bw*n.pid_error + n.new_bw = n.use_bw + cs_junk.K_p*n.use_bw*n.pid_error n.pid_error_sum = n.pid_error n.pid_bw = n.new_bw plog("INFO", "No prev vote for node "+n.nick+": Consensus feedback") @@ -565,8 +598,9 @@ def main(argv): if n.new_bw >= 0xffffffff*1000: plog("WARN", "Bandwidth of node "+n.nick+"="+n.idhex+" exceeded maxint32: "+str(n.new_bw)) n.new_bw = 0xffffffff*1000 - if cs_junk.T_i > 0 and math.fabs(n.pid_error_sum) > \ - math.fabs(2*cs_junk.T_i*n.pid_error/cs_junk.T_i_decay): + if cs_junk.T_i > 0 and cs_junk.T_i_decay > 0 \ + and math.fabs(n.pid_error_sum) > \ + math.fabs(2*cs_junk.T_i*n.pid_error/cs_junk.T_i_decay): plog("NOTICE", "Large pid_error_sum for node "+n.idhex+"="+n.nick+": "+ str(n.pid_error_sum)+" vs "+str(n.pid_error)) if n.new_bw > tot_net_bw*NODE_CAP:

1 0