summaryrefslogtreecommitdiff
Commit message (Collapse)AuthorAge
* Reorder some logs. Squash?pid_logsMike Perry2011-12-15
|
* Clean up some logs. Squash?Mike Perry2011-12-15
|
* Make the guard sample rate a consensus param.Mike Perry2011-12-15
|
* Be more strict w/ circ failures and fast nodes.Mike Perry2011-12-15
| | | | | | | | | | | | | 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.
* Squash.Mike Perry2011-12-13
|
* Use node_class on some other loglines.Mike Perry2011-12-13
|
* Sq..Mike Perry2011-12-13
|
* Use node_class() instead of flags.Mike Perry2011-12-13
|
* SqMike Perry2011-12-13
|
* Log node flags when we cap them.Mike Perry2011-12-13
| | | | Also add some comments.
* Squash.Mike Perry2011-12-12
| | | | (Plz be really close to 0).
* Squash.Mike Perry2011-12-11
|
* Log the network pid error averages.Mike Perry2011-12-11
|
* And it's INT32_MAX, not UINT32_MAX..Mike Perry2011-12-11
| | | | Too bad python doesn't seem to have constants for either?
* Fix an overflow checkMike Perry2011-12-11
| | | | | | | | Tor internally multiplies our values by 1000, and silently eats the overflow. Why do we bother dividing them by 1000 in the first place? It just causes confusion and bugs.. The git blame log and comments mumble something about consensus diffs, but I think we gave up on those...
* Be merciful in Guard intermediate voting mode.Mike Perry2011-12-09
| | | | Also don't use Wgd in mercy mode or in section2 mode.
* And forgot to reset a var...Mike Perry2011-12-05
|
* Fix c_node exception typo.Mike Perry2011-12-05
|
* The old code should use circ failures..loggingMike Perry2011-12-05
| | | | | I think this was added because circ_fail_rate used to get set to 0 by a consensus param that we later changed.
* Always display per-class logs, even if the feature is disabled.Mike Perry2011-12-05
| | | | | I want to try to disable it and see if these become more even or not.
* Improve and demote some noisy loglines.Mike Perry2011-12-04
|
* Clean up the spec a tad.Mike Perry2011-12-03
|
* Allow us to cap feedback via consensus param.Mike Perry2011-12-03
| | | | Also demote a couple logs, and add a couple more.
* Make the pid target bw actually converge.Mike Perry2011-12-02
| | | | | This seems to only take a few seconds to do.. Expensive, but probably not prohibitively so?
* Be more merciful.Mike Perry2011-12-02
| | | | | We can just check to see if we've previously punished the node by inspecting the previous values directly. Forgot about that.
* Implement mercy, subject to consensus parameter.Mike Perry2011-12-02
| | | | | | If the param is set, we don't accumulate negative pid_error_sum for nodes with negative pid_error. Prevents us from punishing these nodes to 0 bw. The param only makes sense for descriptor-based feedback.
* Add a consensus param to raise the PID setpoint.Mike Perry2011-12-02
| | | | | Intuitively, it seems to make sense to exclude nodes with no chance of reaching the setpoint from consideration.
* Use FastFirstHopPK to get circ fail rates for Guards.Mike Perry2011-12-02
| | | | Otherwise we might not detect CPU overload conditions in them.
* Actually, we should be using min(circ_error,pid_error).Mike Perry2011-12-02
|
* Directly use the circ_error if it is negative.Mike Perry2011-12-01
|
* Fix variable shadow, and add some logs.Mike Perry2011-12-01
|
* Change how circuit failure is used.Mike Perry2011-12-01
| | | | Let's try computing a second error and take their sum.
* Implement per-class setpoints.Mike Perry2011-12-01
| | | | Governed by consensus param (off by default).
* Alter default param values.Mike Perry2011-12-01
| | | | | | 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.
* Fix an exception that can happen with very large bandwidths.Mike Perry2011-12-01
| | | | | We have a cap+log warn below that we will hit in this case instead of crashing.
* Use desc bw if consensus bw is missing for some log linesMike Perry2011-11-20
|
* Add some notifications for #4359/#2286.Mike Perry2011-11-20
|
* Be consistent about Wgd application for Guard+Exit nodes.Mike Perry2011-11-20
| | | | Also correct log line.
* Display a total of previous consensus bw measured in INFO logs.Mike Perry2011-11-20
|
* Output git revision.Mike Perry2011-11-20
|
* Bug 2014: Provide url to backup bw server.Mike Perry2011-11-20
|
* Bug #4330: Have snakeinspector attach files if they existsoat-attachmentsMike Perry2011-11-19
|
* Forgot a spec paragraph..Mike Perry2011-11-17
|
* Merge branch 'pid_control-rebased'Mike Perry2011-11-17
|\
| * Let's try turning FetchDirInfo*Early on by default.Mike Perry2011-11-17
| | | | | | | | Memory-only sqlite might make them more bearable.
| * Update spec to include consensus params.Mike Perry2011-11-17
| | | | | | | | | | These are still subject to change, especially in terms of defaults.
| * Allow desc bw to be used in feedback.Mike Perry2011-11-17
| | | | | | | | | | 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.
| * Submodule.Mike Perry2011-11-17
| |
| * Switch to memory-only sqlalchemy by default.Mike Perry2011-11-17
| |
| * Add some comments about potential timeout issues.Mike Perry2011-11-17
| |
generated by cgit v1.2.1 (git 2.18.0) at 2024-02-02 10:44:52 +0000