- tor-commits - lists.torproject.org

[torflow/master] Directly use the circ_error if it is negative.
by mikeperry＠torproject.org 02 Dec '11

02 Dec '11

commit 9e27a77f17bae2f9535373f8b48c2cee6ea2ae4d Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Thu Dec 1 21:23:46 2011 -0800 Directly use the circ_error if it is negative. --- NetworkScanners/BwAuthority/README.spec.txt | 4 +++- NetworkScanners/BwAuthority/aggregate.py | 8 ++++---- 2 files changed, 7 insertions(+), 5 deletions(-) diff --git a/NetworkScanners/BwAuthority/README.spec.txt b/NetworkScanners/BwAuthority/README.spec.txt index db5c125..d223160 100644 --- a/NetworkScanners/BwAuthority/README.spec.txt +++ b/NetworkScanners/BwAuthority/README.spec.txt @@ -516,7 +516,9 @@ to the node, and the average success rate for the entire node class, respectively. - This error value is then added to pid_error. + If this error value is negative (meaning the node is failing + more circuits than the network average), we use it as the + pid_error. "bwauthbestratio=0" If absent, the larger of stream bandwidth vs filtered bandwidth diff --git a/NetworkScanners/BwAuthority/aggregate.py b/NetworkScanners/BwAuthority/aggregate.py index a2372af..ae6f5be 100755 --- a/NetworkScanners/BwAuthority/aggregate.py +++ b/NetworkScanners/BwAuthority/aggregate.py @@ -546,10 +546,10 @@ def main(argv): if cs_junk.use_circ_fails: circ_error = ((1.0-n.circ_fail_rate) - true_circ_avg[n.node_class()]) \ / true_circ_avg[n.node_class()] - # FIXME: Hrmm, should we only penalize for circ successes, or should - # we reward, too? Let's try both for now. - # if circ_error < 0: - n.pid_error += circ_error + # FIXME: Hrmm, should we only penalize for circ fails, or should + # we reward, too? + if circ_error < 0: + n.pid_error = circ_error if n.idhex in prev_votes.vote_map: # If there is a new sample, let's use it for all but guards

1 0

[tor/master] Appease check-spaces
by nickm＠torproject.org 02 Dec '11

02 Dec '11

commit c811b8f3a0a3ec2a071983d6952350f521c8b28d Author: Sebastian Hahn <sebastian(a)torproject.org> Date: Fri Dec 2 06:16:47 2011 +0100 Appease check-spaces --- src/test/test_config.c | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-) diff --git a/src/test/test_config.c b/src/test/test_config.c index 4ce0fa8..d8161de 100644 --- a/src/test/test_config.c +++ b/src/test/test_config.c @@ -156,7 +156,7 @@ test_config_addressmap(void *arg) strlcpy(address, "www.torproject.org", sizeof(address)); test_assert(!addressmap_rewrite(address, sizeof(address), &expires)); -done: + done: ; }

1 0

[tor/master] Work around a false positive in Coverity.
by nickm＠torproject.org 02 Dec '11

02 Dec '11

commit 95af91565bd846be0b199fbc02d7197bf771e867 Author: Sebastian Hahn <sebastian(a)torproject.org> Date: Fri Dec 2 06:15:31 2011 +0100 Work around a false positive in Coverity. Fixes cid 501 and 502. --- src/common/crypto.c | 7 +++++-- 1 files changed, 5 insertions(+), 2 deletions(-) diff --git a/src/common/crypto.c b/src/common/crypto.c index 62b0bce..2aa3fc9 100644 --- a/src/common/crypto.c +++ b/src/common/crypto.c @@ -1912,9 +1912,12 @@ crypto_get_stored_dynamic_dh_modulus(const char *fname) } /* 'fname' contains the DH parameters stored in base64-ed DER - format. We are only interested in the DH modulus. */ + * format. We are only interested in the DH modulus. + * NOTE: We allocate more storage here than we need. Since we're already + * doing that, we can also add 1 byte extra to appease Coverity's + * scanner. */ - cp = base64_decoded_dh = tor_malloc_zero(strlen(contents)); + cp = base64_decoded_dh = tor_malloc_zero(strlen(contents) + 1); length = base64_decode((char *)base64_decoded_dh, strlen(contents), contents, strlen(contents)); if (length < 0) {

1 0

[tor/master] Merge remote-tracking branch 'sebastian/coverity_strlen_fp'
by nickm＠torproject.org 02 Dec '11

02 Dec '11

commit cd10013218a6b1cc42568eb91ed9372e6caddf1f Merge: 5ffa710 95af915 Author: Nick Mathewson <nickm(a)torproject.org> Date: Fri Dec 2 00:24:33 2011 -0500 Merge remote-tracking branch 'sebastian/coverity_strlen_fp' src/common/crypto.c | 7 +++++-- src/test/test_config.c | 2 +- 2 files changed, 6 insertions(+), 3 deletions(-)

1 0

[tor/master] Don't segfault when checking whether a not-yet-used intro point should expire
by nickm＠torproject.org 02 Dec '11

02 Dec '11

commit 5ffa7102c00385445d3855bbc481668edf62d139 Author: Robert Ransom <rransom.8774(a)gmail.com> Date: Thu Dec 1 15:26:45 2011 -0800 Don't segfault when checking whether a not-yet-used intro point should expire Found by katmagic. Bugfix on the #3460 branch, not yet in any release. --- src/or/rendservice.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-) diff --git a/src/or/rendservice.c b/src/or/rendservice.c index a360d5c..177f3bf 100644 --- a/src/or/rendservice.c +++ b/src/or/rendservice.c @@ -1957,7 +1957,8 @@ intro_point_should_expire_now(rend_intro_point_t *intro, return 1; } - if (digestmap_size(intro->accepted_intro_rsa_parts) >= + if (intro->accepted_intro_rsa_parts != NULL && + digestmap_size(intro->accepted_intro_rsa_parts) >= INTRO_POINT_LIFETIME_INTRODUCTIONS) { /* This intro point has been used too many times. Expire it now. */ return 1;

1 0

[torflow/master] Fix variable shadow, and add some logs.
by mikeperry＠torproject.org 02 Dec '11

02 Dec '11

commit c183c9fe3970c9bd022d1d7816860dab62f34d07 Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Thu Dec 1 21:03:06 2011 -0800 Fix variable shadow, and add some logs. --- NetworkScanners/BwAuthority/aggregate.py | 32 +++++++++++++++-------------- 1 files changed, 17 insertions(+), 15 deletions(-) diff --git a/NetworkScanners/BwAuthority/aggregate.py b/NetworkScanners/BwAuthority/aggregate.py index ccf6436..a2372af 100755 --- a/NetworkScanners/BwAuthority/aggregate.py +++ b/NetworkScanners/BwAuthority/aggregate.py @@ -464,32 +464,34 @@ def main(argv): plog("INFO", "PID control enabled") if cs_junk.group_by_class: - for c in ["Guard+Exit", "Guard", "Exit", "Middle"]: - c_nodes = filter(lambda n: n.node_class() == c, nodes.itervalues()) - true_filt_avg[c] = sum(map(lambda n: n.filt_bw, c_nodes))/float(len(c_nodes)) - true_strm_avg[c] = sum(map(lambda n: n.strm_bw, c_nodes))/float(len(c_nodes)) - true_circ_avg[c] = sum(map(lambda n: (1.0-n.circ_fail_rate), + for cl in ["Guard+Exit", "Guard", "Exit", "Middle"]: + c_nodes = filter(lambda n: n.node_class() == cl, nodes.itervalues()) + true_filt_avg[cl] = sum(map(lambda n: n.filt_bw, c_nodes))/float(len(c_nodes)) + true_strm_avg[cl] = sum(map(lambda n: n.strm_bw, c_nodes))/float(len(c_nodes)) + true_circ_avg[cl] = sum(map(lambda n: (1.0-n.circ_fail_rate), c_nodes))/float(len(c_nodes)) - plog("INFO", "Network true_filt_avg["+c+"]: "+str(true_filt_avg[c])) - plog("INFO", "Network true_circ_avg["+c+"]: "+str(true_circ_avg[c])) + plog("INFO", "Network true_filt_avg["+cl+"]: "+str(true_filt_avg[cl])) + plog("INFO", "Network true_circ_avg["+cl+"]: "+str(true_circ_avg[cl])) else: filt_avg = sum(map(lambda n: n.filt_bw, nodes.itervalues()))/float(len(nodes)) strm_avg = sum(map(lambda n: n.strm_bw, nodes.itervalues()))/float(len(nodes)) circ_avg = sum(map(lambda n: (1.0-n.circ_fail_rate), nodes.itervalues()))/float(len(nodes)) - for c in ["Guard+Exit", "Guard", "Exit", "Middle"]: - true_filt_avg[c] = filt_avg - true_strm_avg[c] = strm_avg - true_circ_avg[c] = circ_avg + for cl in ["Guard+Exit", "Guard", "Exit", "Middle"]: + true_filt_avg[cl] = filt_avg + true_strm_avg[cl] = strm_avg + true_circ_avg[cl] = circ_avg + plog("INFO", "Network true_filt_avg: "+str(true_filt_avg["Middle"])) + plog("INFO", "Network true_circ_avg: "+str(true_circ_avg["Middle"])) else: plog("INFO", "PID control disabled") filt_avg = sum(map(lambda n: n.filt_bw*(1.0-n.circ_fail_rate), nodes.itervalues()))/float(len(nodes)) strm_avg = sum(map(lambda n: n.strm_bw*(1.0-n.circ_fail_rate), nodes.itervalues()))/float(len(nodes)) - for c in ["Guard+Exit", "Guard", "Exit", "Middle"]: - true_filt_avg[c] = filt_avg - true_strm_avg[c] = strm_avg + for cl in ["Guard+Exit", "Guard", "Exit", "Middle"]: + true_filt_avg[cl] = filt_avg + true_strm_avg[cl] = strm_avg prev_votes = None @@ -525,7 +527,7 @@ def main(argv): tot_net_bw = 0 for n in nodes.itervalues(): n.fbw_ratio = n.filt_bw/true_filt_avg[n.node_class()] - n.sbw_ratio = n.strm_bw/true_strm_avg[n.node_clasS()] + n.sbw_ratio = n.strm_bw/true_strm_avg[n.node_class()] if cs_junk.bwauth_pid_control: if cs_junk.use_desc_bw:

1 0

[torflow/master] Fix an exception that can happen with very large bandwidths.
by mikeperry＠torproject.org 02 Dec '11

02 Dec '11

commit 325bfcc0f000808c5528f90d69996ab0a186aa10 Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Thu Dec 1 16:56:46 2011 -0800 Fix an exception that can happen with very large bandwidths. We have a cap+log warn below that we will hit in this case instead of crashing. --- NetworkScanners/BwAuthority/aggregate.py | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-) diff --git a/NetworkScanners/BwAuthority/aggregate.py b/NetworkScanners/BwAuthority/aggregate.py index 46dce2e..19697eb 100755 --- a/NetworkScanners/BwAuthority/aggregate.py +++ b/NetworkScanners/BwAuthority/aggregate.py @@ -345,7 +345,7 @@ def main(argv): ns_list = c.get_network_status() for n in ns_list: if n.bandwidth == None: n.bandwidth = -1 - ns_list.sort(lambda x, y: y.bandwidth - x.bandwidth) + ns_list.sort(lambda x, y: int(y.bandwidth/10000.0 - x.bandwidth/10000.0)) for n in ns_list: if n.bandwidth == -1: n.bandwidth = None got_ns_bw = False

1 0

[torflow/master] Alter default param values.
by mikeperry＠torproject.org 02 Dec '11

02 Dec '11

commit 4a4b8a73185f763f0def3e0d30c052f3abeb6fa0 Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Thu Dec 1 17:38:45 2011 -0800 Alter default param values. These defaults should still be equivalent to Section 2 in the spec, but these values will keep the PID interim state from disappearing every time ides crashes and the consensus params disappear. --- NetworkScanners/BwAuthority/README.spec.txt | 73 ++++++++++++++++----------- NetworkScanners/BwAuthority/aggregate.py | 24 ++++---- 2 files changed, 56 insertions(+), 41 deletions(-) diff --git a/NetworkScanners/BwAuthority/README.spec.txt b/NetworkScanners/BwAuthority/README.spec.txt index b8f86d4..81468b5 100644 --- a/NetworkScanners/BwAuthority/README.spec.txt +++ b/NetworkScanners/BwAuthority/README.spec.txt @@ -483,23 +483,59 @@ 3.6. Consensus Parameters The bandwidth auths listen for several consensus parameters to tweak - behavior: + behavior. - "bwauthpid=1" - If present, enables the PID control features in Section 3. + In the absence of any consensus parameters, the default behavior is + to use the PID control code to produce values identical to Section 2, + by using default values of: + + K_p = 1.0, K_i = 0, K_d = 0, as well as bwauthcircs=0. + + This equivalence was proved in Section 3.1, and has been observed + in practice. + + The available consensus parameters are: + + "bwauthpid=0" + If present, entirely disables the PID control features in + Section 3 and computes bandwidths according to Section 2. + + Setting this value to temporarily disable PID feedback is not + recommended, because it causes the PID code to lose interim + recorded state. + + To temporarily disable PID feedback, simply remove all consensus + parameters, and the system will compute Section 2 values while + retaining PID state. "bwauthcircs=1" If present, F_node is multiplied by (1.0 - circ_fail_rate) as described in Section 3.1. - + + "bwauthbestratio=0" + If absent, the larger of stream bandwidth vs filtered bandwidth + is used to compute F_node. + + If present, only filtered stream bandwidth ratios are used. + + "bwauthnsbw=1" + If present, uses consensus bandwidth to determine new bandwidth + values. + + If absent, uses descriptor bandwidth instead of feeding back + PID control values. This may be functionally equivalent to NS + bandwidth so long as T_i is non-zero, because error will get + accumulated in pid_error_sum as opposed to the consensus value + itself. + "bwauthkp=N" - Sets K_p to N/10000.0 + Sets K_p to N/10000.0. If absent, K_p=1.0. "bwauthti=N" - Sets T_i to N/10000.0. For T_i=0, K_i is set to 0. + Sets T_i to N/10000.0. If T_i=0 or absent, K_i is set to 0. "bwauthtd=N" - Sets T_d to N/10000.0. + Sets T_d to N/10000.0. If absent, K_d=0. "bwauthtidecay=N" Sets T_i_decay to N/10000.0. T_i_decay is an parameter @@ -513,25 +549,4 @@ the T_i'th round has experienced a reduction by T_i_decay for the values of T_i that are relevant to us. - If T_i is 0, K_i_decay is set to 0. - - "bwauthbestratio=1" - If present, the larger of stream bandwidth vs filtered bandwidth - is used to compute F_node. - - "bwauthdescbw=1" - If present, uses descriptor bandwidth instead of feeding back - PID control values. This may be functionally equivalent to NS - bandwidth so long as T_i is non-zero, because error will get - accumulated in pid_error_sum as opposed to the consensus value - itself. - - It also has the advantage of allowing the PID control code to be - exercised, yet to still produce identical results to Section 2 - by using the following consensus parameters: - - bwauthpid=1 bwauthdescbw=1 bwauthbestratio=1 bwauthcircs=0 - bwauthkp=10000 bwauthti=0 bwauthtd=0 - - This equivalence was proved in Section 3.1, and has been observed - in practice. + If T_i is 0 or absent, K_i_decay is set to 0. diff --git a/NetworkScanners/BwAuthority/aggregate.py b/NetworkScanners/BwAuthority/aggregate.py index 19697eb..6283f3f 100755 --- a/NetworkScanners/BwAuthority/aggregate.py +++ b/NetworkScanners/BwAuthority/aggregate.py @@ -28,19 +28,19 @@ GUARD_SAMPLE_RATE = 2*7*24*60*60 # 2wks K_p = 1.0 # We expect to correct steady state error in 5 samples (guess) -T_i = 5.0 +T_i = 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 +T_i_decay = 0 # 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 uptake time # vs measurement rate -T_d = 0.5 +T_d = 0 NODE_CAP = 0.05 @@ -221,10 +221,10 @@ class VoteSet: 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.bwauth_pid_control = True self.use_circ_fails = False - self.use_best_ratio = False - self.use_desc_bw = False + self.use_best_ratio = True + self.use_desc_bw = True self.K_p = K_p self.T_i = T_i @@ -235,16 +235,16 @@ class ConsensusJunk: 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 == "bwauthdescbw=1": - self.use_desc_bw = True + if p == "bwauthpid=0": + self.bwauth_pid_control = False + elif p == "bwauthnsbw=1": + self.use_desc_bw = False plog("INFO", "Using descriptor bandwidth") elif p == "bwauthcircs=1": self.use_circ_fails = True plog("INFO", "Counting circuit failures") - elif p == "bwauthbestratio=1": - self.use_best_ratio = True + elif p == "bwauthbestratio=0": + self.use_best_ratio = False plog("INFO", "Choosing larger of sbw vs fbw") elif p.startswith("bwauthkp="): self.K_p = int(p.split("=")[1])/10000.0

1 0

[torflow/master] Implement per-class setpoints.
by mikeperry＠torproject.org 02 Dec '11

02 Dec '11

commit 4f4bc10fba85adc877481e3fb4445847a208da12 Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Thu Dec 1 20:09:25 2011 -0800 Implement per-class setpoints. Governed by consensus param (off by default). --- NetworkScanners/BwAuthority/README.spec.txt | 5 ++ NetworkScanners/BwAuthority/aggregate.py | 65 ++++++++++++++++++++++----- 2 files changed, 58 insertions(+), 12 deletions(-) diff --git a/NetworkScanners/BwAuthority/README.spec.txt b/NetworkScanners/BwAuthority/README.spec.txt index 81468b5..acce106 100644 --- a/NetworkScanners/BwAuthority/README.spec.txt +++ b/NetworkScanners/BwAuthority/README.spec.txt @@ -528,6 +528,11 @@ accumulated in pid_error_sum as opposed to the consensus value itself. + "bwauthbyclass=1" + If present, computes F_avg (see Section 3.1) for each class + of Guard, Middle, Exit, and Guard+Exit nodes, and uses these + flag-specific averages to compute pid_error. + "bwauthkp=N" Sets K_p to N/10000.0. If absent, K_p=1.0. diff --git a/NetworkScanners/BwAuthority/aggregate.py b/NetworkScanners/BwAuthority/aggregate.py index 6283f3f..120de07 100755 --- a/NetworkScanners/BwAuthority/aggregate.py +++ b/NetworkScanners/BwAuthority/aggregate.py @@ -92,6 +92,7 @@ class Node: self.new_bw = None self.change = None self.use_bw = -1 + self.flags = "" # measurement vars from bwauth lines self.measured_at = 0 @@ -132,6 +133,16 @@ class Node: + kd*self.use_bw*self.d_error_dt() return self.pid_bw + def node_class(self): + if "Guard" in self.flags and "Exit" in self.flags: + return "Guard+Exit" + elif "Guard" in self.flags: + return "Guard" + elif "Exit" in self.flags: + return "Exit" + else: + return "Middle" + # Time-weighted sum of error per unit of time (measurement sample) def integral_error(self): if self.prev_error == 0: @@ -222,6 +233,7 @@ class ConsensusJunk: def __init__(self, c): cs_bytes = c.sendAndRecv("GETINFO dir/status-vote/current/consensus\r\n")[0][2] self.bwauth_pid_control = True + self.group_by_class = False self.use_circ_fails = False self.use_best_ratio = True self.use_desc_bw = True @@ -246,6 +258,9 @@ class ConsensusJunk: elif p == "bwauthbestratio=0": self.use_best_ratio = False plog("INFO", "Choosing larger of sbw vs fbw") + elif p == "bwauthbyclass=1": + self.group_by_class = True + plog("INFO", "Grouping nodes by flag-class") elif p.startswith("bwauthkp="): self.K_p = int(p.split("=")[1])/10000.0 plog("INFO", "Got K_p=%f from consensus." % self.K_p) @@ -439,23 +454,45 @@ def main(argv): for n in nodes.itervalues(): n.circ_fail_rate = 0.0 + for idhex in nodes.iterkeys(): + if idhex in prev_consensus: + nodes[idhex].flags = prev_consensus[idhex].flags + + true_filt_avg = {} + true_strm_avg = {} + 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 # socket exhaustion issues.. plog("INFO", "PID control enabled") - true_filt_avg = sum(map(lambda n: n.filt_bw*(1.0-n.circ_fail_rate), - nodes.itervalues()))/float(len(nodes)) - true_strm_avg = sum(map(lambda n: n.strm_bw*(1.0-n.circ_fail_rate), - nodes.itervalues()))/float(len(nodes)) + + if cs_junk.group_by_class: + for c in ["Guard+Exit", "Guard", "Exit", "Middle"]: + c_nodes = filter(lambda n: n.node_class() == c, nodes.itervalues()) + true_filt_avg[c] = sum(map(lambda n: n.filt_bw*(1.0-n.circ_fail_rate), + c_nodes))/float(len(c_nodes)) + true_strm_avg[c] = sum(map(lambda n: n.strm_bw*(1.0-n.circ_fail_rate), + c_nodes))/float(len(c_nodes)) + plog("INFO", "Network true_filt_avg["+c+"]: "+str(true_filt_avg[c])) + else: + filt_avg = sum(map(lambda n: n.filt_bw*(1.0-n.circ_fail_rate), + nodes.itervalues()))/float(len(nodes)) + strm_avg = sum(map(lambda n: n.strm_bw*(1.0-n.circ_fail_rate), + nodes.itervalues()))/float(len(nodes)) + for c in ["Guard+Exit", "Guard", "Exit", "Middle"]: + true_filt_avg[c] = filt_avg + true_strm_avg[c] = strm_avg else: plog("INFO", "PID control disabled") - true_filt_avg = sum(map(lambda n: n.filt_bw, - nodes.itervalues()))/float(len(nodes)) - true_strm_avg = sum(map(lambda n: n.strm_bw, + filt_avg = sum(map(lambda n: n.filt_bw*(1.0-n.circ_fail_rate), + nodes.itervalues()))/float(len(nodes)) + strm_avg = sum(map(lambda n: n.strm_bw*(1.0-n.circ_fail_rate), nodes.itervalues()))/float(len(nodes)) + for c in ["Guard+Exit", "Guard", "Exit", "Middle"]: + true_filt_avg[c] = filt_avg + true_strm_avg[c] = strm_avg - plog("DEBUG", "Network true_filt_avg: "+str(true_filt_avg)) prev_votes = None if cs_junk.bwauth_pid_control: @@ -489,8 +526,8 @@ def main(argv): tot_net_bw = 0 for n in nodes.itervalues(): - n.fbw_ratio = n.filt_bw/true_filt_avg - n.sbw_ratio = n.strm_bw/true_strm_avg + n.fbw_ratio = n.filt_bw/true_filt_avg[n.node_class()] + n.sbw_ratio = n.strm_bw/true_strm_avg[n.node_clasS()] if cs_junk.bwauth_pid_control: if cs_junk.use_desc_bw: @@ -500,9 +537,13 @@ def main(argv): # 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 + n.pid_error = (n.strm_bw*(1.0-n.circ_fail_rate) - + true_strm_avg[n.node_class()]) \ + / true_strm_avg[n.node_class()] else: - n.pid_error = (n.filt_bw*(1.0-n.circ_fail_rate) - true_filt_avg)/true_filt_avg + n.pid_error = (n.filt_bw*(1.0-n.circ_fail_rate) - + true_filt_avg[n.node_class()]) \ + / true_filt_avg[n.node_class()] if n.idhex in prev_votes.vote_map: # If there is a new sample, let's use it for all but guards

1 0

[torflow/master] Change how circuit failure is used.
by mikeperry＠torproject.org 02 Dec '11

02 Dec '11

commit 83f29ce70d0846206daa76f7ad8ab2cf3f5efbd6 Author: Mike Perry <mikeperry-git(a)fscked.org> Date: Thu Dec 1 20:54:46 2011 -0800 Change how circuit failure is used. Let's try computing a second error and take their sum. --- NetworkScanners/BwAuthority/README.spec.txt | 16 +++++++--- NetworkScanners/BwAuthority/aggregate.py | 40 +++++++++++++++------------ 2 files changed, 33 insertions(+), 23 deletions(-) diff --git a/NetworkScanners/BwAuthority/README.spec.txt b/NetworkScanners/BwAuthority/README.spec.txt index acce106..db5c125 100644 --- a/NetworkScanners/BwAuthority/README.spec.txt +++ b/NetworkScanners/BwAuthority/README.spec.txt @@ -369,9 +369,7 @@ 3.1. Modeling Measurement as PID Control The bandwidth authorities measure F_node: the filtered stream - capacity through a given node (filtering is described in Section 1.6) - times the circuit success rate of circuit EXTENDs to the node - (1.0 - circ_fail_rate). + capacity through a given node (filtering is described in Section 1.6). In PID control, we add in this extra failure rate as a damper, to prevent the PID control system from driving nodes to CPU overload. Once nodes @@ -509,8 +507,16 @@ retaining PID state. "bwauthcircs=1" - If present, F_node is multiplied by (1.0 - circ_fail_rate) - as described in Section 3.1. + If present, an additional circ_error value is computed for each + node similar to pid_error of Section 3.1. This value is: + + circ_error = (circ_rate - circ_avg_rate)/circ_avg_rate + + Where circ_rate and circ_avg_rate are the EXTEND success rates + to the node, and the average success rate for the entire node class, + respectively. + + This error value is then added to pid_error. "bwauthbestratio=0" If absent, the larger of stream bandwidth vs filtered bandwidth diff --git a/NetworkScanners/BwAuthority/aggregate.py b/NetworkScanners/BwAuthority/aggregate.py index 120de07..ccf6436 100755 --- a/NetworkScanners/BwAuthority/aggregate.py +++ b/NetworkScanners/BwAuthority/aggregate.py @@ -449,17 +449,13 @@ 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 - for idhex in nodes.iterkeys(): if idhex in prev_consensus: nodes[idhex].flags = prev_consensus[idhex].flags true_filt_avg = {} true_strm_avg = {} + true_circ_avg = {} if cs_junk.bwauth_pid_control: # Penalize nodes for circuit failure: it indicates CPU pressure @@ -470,19 +466,21 @@ def main(argv): if cs_junk.group_by_class: for c in ["Guard+Exit", "Guard", "Exit", "Middle"]: c_nodes = filter(lambda n: n.node_class() == c, nodes.itervalues()) - true_filt_avg[c] = sum(map(lambda n: n.filt_bw*(1.0-n.circ_fail_rate), - c_nodes))/float(len(c_nodes)) - true_strm_avg[c] = sum(map(lambda n: n.strm_bw*(1.0-n.circ_fail_rate), - c_nodes))/float(len(c_nodes)) + true_filt_avg[c] = sum(map(lambda n: n.filt_bw, c_nodes))/float(len(c_nodes)) + true_strm_avg[c] = sum(map(lambda n: n.strm_bw, c_nodes))/float(len(c_nodes)) + true_circ_avg[c] = sum(map(lambda n: (1.0-n.circ_fail_rate), + c_nodes))/float(len(c_nodes)) plog("INFO", "Network true_filt_avg["+c+"]: "+str(true_filt_avg[c])) + plog("INFO", "Network true_circ_avg["+c+"]: "+str(true_circ_avg[c])) else: - filt_avg = sum(map(lambda n: n.filt_bw*(1.0-n.circ_fail_rate), - nodes.itervalues()))/float(len(nodes)) - strm_avg = sum(map(lambda n: n.strm_bw*(1.0-n.circ_fail_rate), - nodes.itervalues()))/float(len(nodes)) + filt_avg = sum(map(lambda n: n.filt_bw, nodes.itervalues()))/float(len(nodes)) + strm_avg = sum(map(lambda n: n.strm_bw, nodes.itervalues()))/float(len(nodes)) + circ_avg = sum(map(lambda n: (1.0-n.circ_fail_rate), + nodes.itervalues()))/float(len(nodes)) for c in ["Guard+Exit", "Guard", "Exit", "Middle"]: true_filt_avg[c] = filt_avg true_strm_avg[c] = strm_avg + true_circ_avg[c] = circ_avg else: plog("INFO", "PID control disabled") filt_avg = sum(map(lambda n: n.filt_bw*(1.0-n.circ_fail_rate), @@ -535,16 +533,22 @@ def main(argv): 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[n.node_class()]) \ + n.pid_error = (n.strm_bw - true_strm_avg[n.node_class()]) \ / true_strm_avg[n.node_class()] else: - n.pid_error = (n.filt_bw*(1.0-n.circ_fail_rate) - - true_filt_avg[n.node_class()]) \ + n.pid_error = (n.filt_bw - true_filt_avg[n.node_class()]) \ / true_filt_avg[n.node_class()] + # Penalize nodes for circ failure rate + if cs_junk.use_circ_fails: + circ_error = ((1.0-n.circ_fail_rate) - true_circ_avg[n.node_class()]) \ + / true_circ_avg[n.node_class()] + # FIXME: Hrmm, should we only penalize for circ successes, or should + # we reward, too? Let's try both for now. + # if circ_error < 0: + n.pid_error += circ_error + if n.idhex in prev_votes.vote_map: # If there is a new sample, let's use it for all but guards if n.measured_at > prev_votes.vote_map[n.idhex].measured_at:

1 0