Hi All,

I have revised proposal #210 - Faster Headless Consensus Bootstrapping today, after a number of discussions with Peter Palfrader  Nick Mathewson, Mike Perry, and others.

This proposal aims to improve tor’s consensus download behaviour when the authorities (or directory mirrors) are down. It has tor initiate multiple concurrent consensus connections, then download the consensus through the first TLS connection that completes.

This proposal is a solution to bug #4483 - If k of n authorities are down, k/n bootstrapping clients are delayed for minutes.
It is also needed to implement #15775 - Add IPv4 Fallback Directory List to tor... and #8374 - Ship list of fallback directory mirrors on long-term fixed IPv6 addresses.

The key changes are:
* modify the scheme to perform exponential backoff on connections, rather than connections in batches
* modify the scheme to enable IPv6 bootstrap on IPv6-only clients (see also #17217 - Change clients to automatically use IPv6 if they can bootstrap over it)
* specify a way that clients can still benefit from clock verification via TLS connections with the authorities, without downloading the entire consensus from the authorities if it is available sooner from a mirror
* analyse the expected failure rate and additional connection load imposed by this proposal

The full text is included below, and a branch with these changes is available as bootstrap-exponential-backoff in https://github.com/teor2345/torspec.git

Please feel free to respond here, or on the #4483 ticket at
https://trac.torproject.org/projects/tor/ticket/4483

Thanks

Tim

-----

Filename: 210-faster-headless-consensus-bootstrap.txt
Title: Faster Headless Consensus Bootstrapping
Author: Mike Perry, Tim Wilson-Brown, Peter Palfrader
Created: 01-10-2012
Last Modified: 02-10-2015
Status: Open
Target: 0.2.8.x+


Overview and Motiviation

 This proposal describes a way for clients to fetch the initial
 consensus more quickly in situations where some or all of the directory
 authorities are unreachable. This proposal is meant to describe a
 solution for bug #4483.

Design: Bootstrap Process Changes

 The core idea is to attempt to establish bootstrap connections in
 parallel during the bootstrap process, and download the consensus from
 the first connection that completes.

 Connection attempts will be performed on an exponential backoff basis.
 Initially, connections will be performed to a randomly chosen hard
 coded directory mirror and a randomly chosen canonical directory
 authority. If neither of these connections complete, additional mirror
 and authority connections are tried. Mirror connections are tried at
 a faster rate than authority connections.

 We specify that mirror connections retry after half a second, and then
 double the retry time with every connection:
 0, 1, 2, 4, 8, 16, 32, ...

 We specify that directory authority connections retry after 5 seconds,
 and then double the retry time with every connection:
 0, 10, 20, ...

 If the client has both an IPv4 and IPv6 address, we try IPv4 and IPv6
 mirrors and authorities on the following schedule:
 IPv4, IPv6, IPv4, IPv6, ...

 We try IPv4 first to avoid overloading IPv6-enabled authorities and
 mirrors. Mirrors and auths get a separate IPv4/IPv6 schedule. This
 ensures that we try an IPv6 authority within the first 10 seconds.
 This helps implement #8374 and related tickets.

 The maximum retry time for both timers is 3 days + 1 hour. This places a
 small load on the mirrors and authorities, while allowing a client that
 regains a network connection to eventually download a consensus.

 The retry timers must reset on HUP and any network reachability events,
 [ TODO: do we have network reachability events? ]
 so that clients that have unreliable networks can recover from network
 failures.

 The first connection to complete will be used to download the consensus
 document and the others will be closed, after which bootstrapping will
 proceed as normal.

 A benefit of connecting to directory authorities is that clients are
 warned if their clock is wrong. Therefore, when closing a directory
 authority connection, we check to see if we have successfully connected
 to an authority during this run of the Tor client. If not, we allow the
 authority TLS connection to complete, then close the connection.

 We expect the vast majority of clients to succeed within 4 seconds,
 after making up to 4 connection attempts to mirrors and 1 connection
 attempt to an authority. Clients which can't connect in the first
 10 seconds, will try 1 more mirror, then try to contact another
 directory authority. We expect almost all clients to succeed within
 10 seconds. This is a much better success rate than the current Tor
 implementation, which fails k/n of clients if k of the n directory
 authorities are down. (Or, if the connection fails in certain ways,
 (k/n)^2.)

 If at any time, the total outstanding bootstrap connection attempts
 exceeds 10, no new connection attempts are to be launched until an
 existing connection attempt experiences full timeout. The retry time
 is not doubled when a connection is skipped.

Design: Fallback Dir Mirror Selection

 The set of hard coded directory mirrors from #572 shall be chosen using
 the 100 Guard nodes with the longest uptime.

 The fallback weights will be set using each mirror's fraction of
 consensus bandwidth out of the total of all 100 mirrors, adjusted to
 ensure no fallback directory sees more than 10% of clients. We will
 also exclude fallback directories that are less than 1/1000 of the
 consensus weight, as they are not large enough to make it worthwhile
 including them.

 This list of fallback dir mirrors should be updated with every
 major Tor release. In future releases, the number of dir mirrors
 should be set at 20% of the current Guard nodes (approximately 200 as
 of October 2015), rather than fixed at 100.

Performance: Additional Load with Current Parameter Choices

 This design and the connection count parameters were chosen such that
 no additional bandwidth load would be placed on the directory
 authorities. In fact, the directory authorities should experience less
 load, because they will not need to serve the consensus document for a
 connection in the event that one of the directory mirrors complete their
 connection before the directory authority does.

 However, the scheme does place additional TLS connection load on the
 fallback dir mirrors. Because bootstrapping is rare and all but one of 
 the TLS connections will be very short-lived and unused, this should not
 be a substantial issue.

 The dangerous case is in the event of a prolonged consensus failure
 that induces all clients to enter into the bootstrap process. In this
 case, the number of TLS connections to the fallback dir mirrors within
 the first second would be 2*C/100, or 40,000 for C=2,000,000 users. If
 no connections complete before the 10 retries, 7 of which go to
 mirrors, this could reach as high as 140,000 connection attempts, but
 this is extremely unlikely to happen in full aggregate.

 However, in the no-consensus scenario today, the directory authorities
 would already experience 2*C/9 or 444,444 connection attempts. (Tor
 currently tries 2 authorities, before delaying the next attempt.) The
 10-retry scheme, 3 of which go to authorities, increases their total
 maximum load to about  666,666 connection attempts, but again this is
 unlikely to be reached in aggregate. Additionally, with this scheme,
 even if the dirauths are taken down by this load, the dir mirrors
 should be able to survive it.

Implementation Notes: Code Modifications

 The implementation of the bootstrap process is unfortunately mixed 
 in with many types of directory activity.

 The process starts in update_consensus_networkstatus_downloads(),
 which initiates a single directory connection through
 directory_get_from_dirserver(). Depending on bootstrap state,
 a single directory server is selected and a connection is
 eventually made through directory_initiate_command_rend().

 There appear to be a few options for altering this code to retry multiple
 simultaneous connections. Without refactoring, one approach would be to
 set a connection retry helper function timer in
 directory_initiate_command_routerstatus() from
 directory_get_from_dirserver() if the purpose is
 DIR_PURPOSE_FETCH_CONSENSUS and the only directory servers available
 are the authorities and the fallback dir mirrors. (That is, there is no
 valid consensus.)  The retry helper function would check the list of
 pending connections and, if it is 10 or greater, skip the connection
 attempt, and leave the retry time constant.

 The code in directory_initiate_command_rend() would then need to be
 altered to maintain a list of the dircons created for this purpose as
 well as avoid immediately queuing the directory_send_command() request
 for the DIR_PURPOSE_FETCH_CONSENSUS purpose. A flag would need to be set
 on the dircon to be checked in connection_dir_finished_connecting().

 The function connection_dir_finished_connecting() would need to be
 altered to examine the list of pending dircons, determine if this one is
 the first to complete, and if so, then call directory_send_command() to
 download the consensus and close the other pending dircons.
 connection_dir_finished_connecting() would also cancel the timer.

Reliability Analysis

 We make the pessimistic assumptions that 50% of connections to directory
 mirrors fail, and that 20% of connections to authorities fail. (Actual
 figures depend on relay churn, age of the fallback list, and authority
 uptime.)

 We expect the first 10 connection retry times to be:
 Mirror:   0s  1s  2s    4s    8s           16s             32s
 Auth:     0s                        10s            20s
 Success: 90% 95% 97% 98.7% 99.4% 99.89% 99.94% 99.988% 99.994%

 97%    of clients succeed in the first 2 seconds.
 99.4%  of clients succeed without trying a second authority.
 99.89% of clients succeed in the first 10 seconds.
  0.11% of clients remain, but in this scenario, 2 authorities are down,
        so the client is most likely blocked from the Tor network.

 The current implementation makes 1 or 2 authority connections within the
 first second, depending on exactly how the first connection fails. Under
 the 20% authority failure assumption, these clients would have a success
 rate of either 80% or 96% within a few seconds. The scheme above has a
 greater success rate in the first few seconds, while spreading the load
 among a larger number of directory mirrors. In addition, if all the
 authorities are blocked, current clients will inevitably fail, as they
 do not have a list of directory mirrors.