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.
Clients represent the majority of the load on the network. They can use
directory mirrors to download their documents, as the mirrors download
their documents from the authorities early in the consensus validity
period.
We specify that client mirror connections retry after one second, and
then double the retry time with every connection attempt:
0, 1, 2, 4, 8, 16, 32, ...
(The timers currently implemented in Tor increment with every
connection failure.)
We specify that client directory authority connections retry after
10 seconds, and then double the retry time with every connection:
0, 10, 20, ...
If a client has both an IPv4 and IPv6 address, it will try IPv4 and
IPv6 mirrors and authorities on the following schedule:
IPv4, IPv6, IPv4, IPv6, ...
[ TODO: should we add random noise to these scheduled times? - teor
Tor doesn’t add random noise to the current failure-based
timers, but as failures are a network event, they are
somewhat random/arbitrary already. These attempt-based timers
will go off every few seconds, exactly erraon the second. ]
(Relays can’t use directory mirrors to download their documents,
as they *are* the directory mirrors.)
The maximum retry time for all these 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.
We try IPv4 first to avoid overloading IPv6-enabled authorities and
mirrors. Each timing schedule uses a separate IPv4/IPv6 schedule.
This ensures that clients try an IPv6 authority within the first
10 seconds. This helps implement #8374 and related tickets.
We don't want to keep on trying an IP version that always fails.
Therefore, once sufficient IPv4 and IPv6 connections have been
attempted, we select an IP version for new connections based on the ratio
of their failure rates, up to a maximum of 1:5. This may not make a
substantial difference to consensus downloads, as we only need one
successful consensus download to bootstrap. However, it is important for
future features like #17217, where clients try to automatically determine
if they can use IPv4 or IPv6 to contact the Tor network.
The retry timers and IP version schedules must reset on HUP and any
network reachability events, so that clients that have unreliable networks
can recover from network failures.
[ TODO: Do we do this for any other timers?
I think this needs another proposal, it’s out of scope here.
- teor ]
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.
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,
it will retry once, failing 1-(1-(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.
A benefit of connecting to directory authorities is that clients are
warned if their clock is wrong. Starting the authority and fallback
schedules at the same time should ensure that some clients check their
clock with an authority at each bootstrap.
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.
[TODO: change the script to dynamically calculate an upper limit.]
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 entire 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. It looks like we can modify
update_consensus_networkstatus_downloads() to make connections more often
if the purpose is DIR_PURPOSE_FETCH_CONSENSUS and there is no valid
(reasonably live) consensus. We can make multiple connections from
update_consensus_networkstatus_downloads(), as the sockets are non-blocking.
(This socket appears to be non-blocking on Unixes (SOCK_NONBLOCK & O_NONBLOCK)
and Windows (FIONBIO).) As long as we can tolerate a timer resolution of
~1 second (due to the use of second_elapsed_callback and time_t), this
requires no additional timers or callbacks. We can make 1 connection for each
schedule per second, for a maximum of 2 per second.
The schedules can be specified in:
TestingClientBootstrapConsensusAuthorityDownloadSchedule
TestingClientBootstrapConsensusFallbackDownloadSchedule
(Similar to the existing TestingClientConsensusDownloadSchedule.)
TestingServerIPVersionPreferenceSchedule
(Consisting of a CSV like “4,6,4,6”, or perhaps “0,1,0,1”.)
update_consensus_networkstatus_downloads() checks 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_send_command() and connection_finished_connecting()
would need to be altered to check that we are not already downloading the
consensus. If we’re not, then download the consensus on this connection, and
close any other pending consensus dircons.
We might also need to make similar changes in authority_certs_fetch_missing(),
as we can’t use a consensus until we have enough authority certificates.
However, Tor already makes multiple requests (one per certificate), and only
needs a majority of certificates to validate a consensus. Therefore, we will
only need to modify authority_certs_fetch_missing() if clients download a
consensus, then end up getting stuck downloading certificates. (Current tests
show bootstrapping working well without any changes to authority certificate
fetches.)
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:
(Research shows users tend to lose interest after 40 seconds.)
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
unreachable, so the client is most likely blocked from the Tor
network. Alternately, they will likely succeed on relaunch.
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.