commit 1e6a4219b57b8ea91418bd1c580a5b255e10a491 Author: Mike Perry <mikeperry-git@fscked.org> Date: Thu Dec 15 11:50:42 2011 -0800 Be more strict w/ circ failures and fast nodes. We use unfiltered bandwidth because I suspect we're being too generous to fast nodes. To get more feedback, they should have to have spare capacity all the time, not just sometimes. However, gimpy nodes will still get to use filtered bandwidth if we've punished them too much already. Also, circuit failures now target a 0% rate (full success) instead of the network average, but the penalty only applies if you fall below the network average for your node class. --- NetworkScanners/BwAuthority/aggregate.py | 32 +++++++++++++++++++++-------- 1 files changed, 23 insertions(+), 9 deletions(-) diff --git a/NetworkScanners/BwAuthority/aggregate.py b/NetworkScanners/BwAuthority/aggregate.py index b8b01ea..46e4679 100755 --- a/NetworkScanners/BwAuthority/aggregate.py +++ b/NetworkScanners/BwAuthority/aggregate.py @@ -21,6 +21,7 @@ IGNORE_GUARDS = 0 # The guard measurement period is based on the client turnover # rate for guard nodes +# XXX: Make this a consensus param GUARD_SAMPLE_RATE = 2*7*24*60*60 # 2wks # PID constant defaults. May be overridden by consensus @@ -579,7 +580,17 @@ def main(argv): n.use_bw = n.ns_bw if cs_junk.use_pid_tgt: - n.pid_error = (n.filt_bw - pid_tgt_avg[n.node_class()]) \ + n.pid_error = (n.strm_bw - pid_tgt_avg[n.node_class()]) \ + / pid_tgt_avg[n.node_class()] + # use filt_bw for pid_error < 0 + if cs_junk.use_mercy: + if cs_junk.use_desc_bw: + if n.pid_error_sum < 0 and n.pid_error < 0: + n.pid_error = (n.filt_bw - pid_tgt_avg[n.node_class()]) \ + / pid_tgt_avg[n.node_class()] + else: + if n.desc_bw > n.ns_bw and n.pid_error < 0: + n.pid_error = (n.filt_bw - pid_tgt_avg[n.node_class()]) \ / pid_tgt_avg[n.node_class()] else: if cs_junk.use_best_ratio and n.sbw_ratio > n.fbw_ratio: @@ -593,14 +604,17 @@ def main(argv): # only in the event of update? # Penalize nodes for circ failure rate if cs_junk.use_circ_fails: - # FIXME: Should we compute this relative to 0? Why target anything - # less? - 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 fails, or should - # we reward, too? - if circ_error < 0: - n.pid_error = min(circ_error,n.pid_error) + # Compute circ_error relative to 1.0 (full success), but only + # apply it if it is both below the network avg and worse than + # the pid_error + if (1.0-n.circ_fail_rate) < true_circ_avg[n.node_class()]: + circ_error = -n.circ_fail_rate # ((1.0-fail) - 1.0)/1.0 + if circ_error < 0 and circ_error < n.pid_error: + plog("INFO", + "CPU overload for %s node %s=%s desc=%d ns=%d pid_error=%f circ_error=%f circ_fail=%f" % + (n.node_class(), n.nick, n.idhex, n.desc_bw, n.ns_bw, + n.pid_error, circ_error, n.circ_fail_rate)) + n.pid_error = min(circ_error,n.pid_error) # Don't accumulate too much amplification for fast nodes if cs_junk.use_desc_bw: