root/dotorg/trunk/html/beps/bep_0022.rst

BEP: 22
Title: BitTorrent Local Tracker Discovery Protocol
Version: $Revision$
Last-Modified: $Date$
Author:  David Harrison <dave@bittorrent.com>, Stanislav Shalunov <shalunov@bittorrent.com>, Greg Hazel <greg@bittorrent.com>
Status:  Draft
Type:    Standards track
Content-Type: text/x-rst
Created: 12-May-2008
Post-History: 

Motivation
==========

Some Internet Service Providers (ISPs) may wish to localize traffic 
to reduce transit costs, reduce internal traffic, and improve user
experience by speeding up downloads.  

With this extension, BitTorrent clients are able to discover a tracker
nearby on the network, and via this tracker discover nearby caches or
peers.  A cache may simply be a fast peer in the middle of the network. It
might also have substantial disk space. The client communicates with a
cache using the normal BitTorrent protocol.

When a cache is present, the user benefits from having a high capacity
peer from which the user's client downloads and to which it can
delegate seeding.  When a cache inside the user's ISP network seeds on
behalf of the client, it frees upstream capacity in the user's access
network benefiting the user and those that share the access network.
When subsequent peers transfer from their ISP's cache, the ISP
experiences less transit traffic.

The scope of this BEP is limited to the local tracker discovery
process.  Extensions to the BitTorrent protocol suite to delegate
seeding or improve cache performance are beyond the scope of this BEP.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL
NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED",  "MAY", and
"OPTIONAL" in this document are to be interpreted as described in
IETF RFC 2119 [#RFC-2119]_.
 
Client implementation of local tracker discovery is OPTIONAL.  Clients
MUST NOT announce private torrents to a local tracker.  It is
RECOMMENDED that clients provide a user option for turning off local
tracker discovery.  Local tracker discovery MAY be off by default.
Clients MAY automatically turn off caching if the performance benefit
is not obvious.  Determining obviousness is beyond the scope of this
BEP.

The Discovery Mechanism
=======================

To find the tracker for its ISP, a BitTorrent client performs a reverse
DNS lookup on its external IP address and then obtains the BitTorrent
SRV resource record associated with the host's domain name.  For
example, a host with address 69.107.0.14 obtains the PTR record at

::

  14.0.107.69.in-addr.arpa

The client's host IP address may not match the host's IP address as
seen outside the client's private network.  We address this in Section
`Network Address Translators`_.

The PTR resource record returned for this example contains domain name

::

  adsl-69-107-0-14.dsl.pltn13.pacbell.net

The client then looks up the SRV records at

::
 
  _bittorrent-tracker._tcp.adsl-69-107-0-14.dsl.pltn13.pacbell.net

If no SRV record is found, one or more subsequent queries take place as
described in `Iterative Queries`_.

The target field in each returned SRV resource record contains the
domain name of a tracker and the port on which the tracker runs.  This
tracker is called a *local tracker*, but the protocol to talk to this
tracker is no different from the standard BitTorrent tracker protocol
described in [#BEP-3]_.

When the BitTorrent client joins a swarm it announces to one or more
of the trackers referenced in the .torrent file and announces to the
local tracker.  The local tracker returns peers which may be caches or
other peers that announced the same file to the local tracker.

A client MAY treat nearby peers or caches preferentially.
 
Reverse DNS lookups are described in RFC 1034 [#RFC-1034]_.
The SRV resource record type is described in RFC 2782 [#RFC-2782]_.  


Iterative Queries
=================

The domain name returned from the reverse DNS lookup is specific to
the querying host.  In the naive implementation in DNS, there would be
one SRV resource record for every querying host.  This would work but
is burdensome.  A natural, seemingly less burdensome, but incorrect
solution is to use a wildcard of the form::

  *.pacbell.net

If wildcards are implemented according to the algorithm in section
4.3.2 in [#RFC-1034]_ then all subdomains of pacbell.net that do not
have an exact label match will match the wildcard.  Thus unless there
is an exact match then queries for

::
 
  _bittorrent-tracker._tcp.adsl-69-107-0-14.dsl.pltn13.pacbell.net

and

::

  _jabber._tcp.pacbell.net 

both match \*.pacbell.net and all SRV resource records with owner
\*.pacbell.net would be returned with the name set to the name in the
query.  Thus it would be impossible to disambiguate Jabber from
BitTorrent SRV records without further information.  This behavior is
implemented with BIND 9.4.1.

Another natural but incorrect solution is to specify domain names of
the type

::

  _bittorrent-tracker._tcp.*.pacbell.net

Section 4.3.3 in [#RFC-1034]_ specifies that wildcards only appear as
the first label in a domain name.  This restriction was lifted in
[#RFC-4592]_, but not with semantics applicable to our use case.  An
asterisk not at the beginning of a domain name is not treated like a
wildcard.  Only a lookup for the exact domain name

::

  _bittorrent-tracker._tcp.*.pacbell.net

matches.

We propose an alternative that avoids wildcards and allows
suborganizations to override SRV records provided by parent
organizations: the peer starts by querying using its fully-qualified
domain name returned from the reverse DNS lookup, and if this fails
then it queries again after removing the most specific (leftmost)
label in the domain name.  For example, if no SRV records are returned
when querying for

::

  _bittorrent-tracker._tcp.adsl-69-107-0-14.dsl.pltn13.pacbell.net

then the client queries for

::

  _bittorrent-tracker._tcp.dsl.pltn13.pacbell.net

and then

::

  _bittorrent-tracker._tcp.pltn13.pacbell.net

The search removes one label at a time terminating when one or more
resource records are found or before querying the root domain or
top-level domains that are not ccTLDs, e.g., .com, .org, .net. We
avoid querying the root or top-level domains given the low likelihood
that caches would be defined globally, and thus clients would
unnecessarily burden the root domain name servers with queries
generating negative results. We considered stopping before querying
country-level domains, but a country providing public infrastructure
might choose to provide caches.


Network Address Translators
===========================

Many hosts on the Internet sit in private networks that connect to the
Internet via a Network Address Translator (NAT).  Such hosts may have
an IP address allocated from one of the private IP address ranges
defined by IANA, e.g., ranges with prefixes 10/8, 172.16/12, and
192.168/16.  When communicating with hosts outside the private
network, the NAT translates the private IP to a globally-routable IP
address.  This globally-routable address is the host's *external IP
address*.

The BitTorrent client must use its host's external IP address.  A
BitTorrent client MAY obtain its host's external IP either from the
*external ip* key returned from a tracker implementing BEP 24
[#BEP-24]_ or from peers implementing the *yourip* extension defined
for the *Extension Protocol* proposed in [#BEP-10]_.

Example
=======

In our example, we use AT&T's PacBell network.  AT&T could implement
tracker discovery by adding the following lines to the zone file for
pacbell.net,

::

  ; name                                ttl  cls rr  pri weight port target
  _bittorrent-tracker._tcp.pacbell.net. 600  IN  SRV 5   0      6969 tracker

Now when a client performs tracker discovery, it performs three DNS
queries removing labels before reaching the domain name pacbell.net,
at which point the SRV record is returned and the client queries
tracker.pacbell.net to obtain the domain names of caches.

In Python, the local tracker's port and domain can be obtained using
PyDNS using the following code::

  import DNS

  tlds = ["com", "net", "org"]  # add more TLDs here.
  
  name = DNS.revlookup( "69.107.0.14" )
  names = name.split('.')
  while names and names[0] not in tlds:
     name = "_bittorrent-tracker._tcp." + ".".join(names)
     req = DNS.Request( name=name, qtype="SRV", protocol="udp")
     response = req.req()
     if response.answers:
        break
     del names[0]

  print "response=", response.show()

which might generate output like

::

  response= ; <<>> PDG.py 1.0 <<>> _bittorrent._tcp.pacbell.net SRV
  ;; options: recurs 
  ;; got answer:
  ;; ->>HEADER<<- opcode 0, status NOERROR, id 0
  ;; flags: qr aa rd ra; Ques: 1, Ans: 1, Auth: 2, Addit: 3
  ;; QUESTIONS:
  ;;      _bittorrent-tracker._tcp.pacbell.net, type = SRV, class = IN
  
  ;; ANSWERS:
  _bittorrent-tracker._tcp.pacbell.net    600    SRV     (5, 0, 6969, 'cache.pacbell.net')
  
  ;; AUTHORITY RECORDS:
  pacbell.net             86400   NS      ns1.pbi.net
  pacbell.net             86400   NS      ns2.pbi.net
  
  ;; ADDITIONAL RECORDS:
  cache.pacbell.net       86400   A       69.107.0.1
  ns1.pacbell.net         86400   A       206.13.28.11
  ns2.pacbell.net         86400   A       206.13.29.11
  
  ;; Total query time: 0 msec
  ;; To SERVER: localhost
  ;; WHEN: Mon May 19 16:00:12 2008

The answer above is fictional since AT&T does not at this time
implement SRV records for BitTorrent trackers.

In Microsoft Windows, the port and domain name of the server can be
obtained using WinDNS (Dnsapi.lib) using DnsQuery().  In Unix, the
relevant call is res_query() from libresolv.

References
==========

.. [#BEP-3] BEP_0003. The BitTorrent Protocol Specification, Cohen.
   http://www.bittorrent.org/beps/bep_0003.html

.. [#BEP-10] BEP_0010.  Extension Protocol. Norberg, Strigeus, Hazel.
   http://www.bittorrent.org/beps/bep_0010.html

.. [#BEP-24] BEP_0024.  Tracker Returns External IP.  Harrison.
   http://www.bittorrent.org/beps/bep_0024.html

.. [#RFC-1034] RFC-1034.  DOMAIN NAMES - CONCEPTS AND FACILITIES. Mockapetris,
   November 1987. http://tools.ietf.org/html/rfc1034

.. [#RFC-2119] RFC-2119. http://www.ietf.org/rfc/rfc2119.txt

.. [#RFC-2782] RFC-2782.  A DNS RR for specifying the location of services (DNS
   SRV). Gulbrandsen, Vixie, Esibov. February 2000. 
   http://tools.ietf.org/html/rfc2782

.. [#RFC-4592] RFC-4592. The Role of Wildcards in the Domain Name
   System. Lewis.  http://www.faqs.org/rfcs/rfc4592.html




Copyright
=========

This document has been placed in the public domain.



..
   Local Variables:
   mode: indented-text
   indent-tabs-mode: nil
   sentence-end-double-space: t
   fill-column: 70
   coding: utf-8
   End: