RFC 7363






Internet Engineering Task Force (IETF)                        J. Maenpaa
Request for Comments: 7363                                  G. Camarillo
Category: Standards Track                                       Ericsson
ISSN: 2070-1721                                           September 2014


                Self-Tuning Distributed Hash Table (DHT)
              for REsource LOcation And Discovery (RELOAD)

Abstract



   REsource LOcation And Discovery (RELOAD) is a peer-to-peer (P2P)
   signaling protocol that provides an overlay network service.  Peers
   in a RELOAD overlay network collectively run an overlay algorithm to
   organize the overlay and to store and retrieve data.  This document
   describes how the default topology plugin of RELOAD can be extended
   to support self-tuning, that is, to adapt to changing operating
   conditions such as churn and network size.

Status of This Memo



   This is an Internet Standards Track document.

   This document is a product of the Internet Engineering Task Force
   (IETF).  It represents the consensus of the IETF community.  It has
   received public review and has been approved for publication by the
   Internet Engineering Steering Group (IESG).  Further information on
   Internet Standards is available in Section 2 of RFC 5741.

   Information about the current status of this document, any errata,
   and how to provide feedback on it may be obtained at
   http://www.rfc-editor.org/info/rfc7363.

Copyright Notice



   Copyright (c) 2014 IETF Trust and the persons identified as the
   document authors.  All rights reserved.

   This document is subject to BCP 78 and the IETF Trust's Legal
   Provisions Relating to IETF Documents
   (http://trustee.ietf.org/license-info) in effect on the date of
   publication of this document.  Please review these documents
   carefully, as they describe your rights and restrictions with respect
   to this document.  Code Components extracted from this document must
   include Simplified BSD License text as described in Section 4.e of
   the Trust Legal Provisions and are provided without warranty as
   described in the Simplified BSD License.




Maenpaa & Camarillo          Standards Track                    [Page 1]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


Table of Contents



   1. Introduction ....................................................2
   2. Terminology .....................................................3
   3. Introduction to Stabilization in DHTs ...........................5
      3.1. Reactive versus Periodic Stabilization .....................5
      3.2. Configuring Periodic Stabilization .........................6
      3.3. Adaptive Stabilization .....................................7
   4. Introduction to Chord ...........................................7
   5. Extending Chord-Reload to Support Self-Tuning ...................9
      5.1. Update Requests ............................................9
      5.2. Neighbor Stabilization ....................................10
      5.3. Finger Stabilization ......................................11
      5.4. Adjusting Finger Table Size ...............................11
      5.5. Detecting Partitioning ....................................11
      5.6. Leaving the Overlay .......................................11
   6. Self-Tuning Chord Parameters ...................................12
      6.1. Estimating Overlay Size ...................................12
      6.2. Determining Routing Table Size ............................13
      6.3. Estimating Failure Rate ...................................13
           6.3.1. Detecting Failures .................................14
      6.4. Estimating Join Rate ......................................14
      6.5. Estimate Sharing ..........................................15
      6.6. Calculating the Stabilization Interval ....................17
   7. Overlay Configuration Document Extension .......................17
   8. Security Considerations ........................................18
   9. IANA Considerations ............................................18
      9.1. Message Extensions ........................................18
      9.2. New Overlay Algorithm Type ................................19
      9.3. A New IETF XML Registry ...................................19
   10. Acknowledgments ...............................................19
   11. References ....................................................19
      11.1. Normative References .....................................19
      11.2. Informative References ...................................20

1.  Introduction



   REsource LOcation And Discovery (RELOAD) [RFC6940] is a peer-to-peer
   signaling protocol that can be used to maintain an overlay network
   and to store data in and retrieve data from the overlay.  For
   interoperability reasons, RELOAD specifies one overlay algorithm,
   called "chord-reload", that is mandatory to implement.  This document
   extends the chord-reload algorithm by introducing self-tuning
   behavior.

   DHT-based overlay networks are self-organizing, scalable, and
   reliable.  However, these features come at a cost: peers in the
   overlay network need to consume network bandwidth to maintain routing



Maenpaa & Camarillo          Standards Track                    [Page 2]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


   state.  Most DHTs use a periodic stabilization routine to counter the
   undesirable effects of churn on routing.  To configure the parameters
   of a DHT, some characteristics such as churn rate and network size
   need to be known in advance.  These characteristics are then used to
   configure the DHT in a static fashion by using fixed values for
   parameters such as the size of the successor set, size of the routing
   table, and rate of maintenance messages.  The problem with this
   approach is that it is not possible to achieve a low failure rate and
   a low communication overhead by using fixed parameters.  Instead, a
   better approach is to allow the system to take into account the
   evolution of network conditions and adapt to them.

   This document extends the mandatory-to-implement chord-reload
   algorithm by making it self-tuning.  The use of the self-tuning
   feature is optional.  However, when used, it needs to be supported by
   all peers in the RELOAD overlay network.  The fact that a RELOAD
   overlay uses the self-tuning feature is indicated in the RELOAD
   overlay configuration document using the CHORD-SELF-TUNING algorithm
   name specified in Section 9.2 in the topology-plugin element.  Two
   main advantages of self-tuning are that users no longer need to tune
   every DHT parameter correctly for a given operating environment and
   that the system adapts to changing operating conditions.

   The remainder of this document is structured as follows: Section 2
   provides definitions of terms used in this document.  Section 3
   discusses alternative approaches to stabilization operations in DHTs,
   including reactive stabilization, periodic stabilization, and
   adaptive stabilization.  Section 4 gives an introduction to the Chord
   DHT algorithm.  Section 5 describes how this document extends the
   stabilization routine of the chord-reload algorithm.  Section 6
   describes how the stabilization rate and routing table size are
   calculated in an adaptive fashion.

2.  Terminology



   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
   "OPTIONAL" in this document are to be interpreted as described in
   [RFC2119].

   This document uses terminology and definitions from the RELOAD base
   specification [RFC6940].

   numBitsInNodeId:  Specifies the number of bits in a RELOAD Node-ID.

   DHT:  Distributed Hash Tables are a class of decentralized
      distributed systems that provide a lookup service similar to a
      regular hash table.  Given a key, any peer participating in the



Maenpaa & Camarillo          Standards Track                    [Page 3]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


      system can retrieve the value associated with that key.  The
      responsibility for maintaining the mapping from keys to values is
      distributed among the peers.

   Chord Ring:  The Chord DHT uses ring topology and orders identifiers
      on an identifier circle of size 2^numBitsInNodeId.  This
      identifier circle is called the Chord ring.  On the Chord ring,
      the responsibility for a key k is assigned to the node whose
      identifier equals to or immediately follows k.

   Finger Table:  A data structure with up to (but typically less than)
      numBitsInNodeId entries maintained by each peer in a Chord-based
      overlay.  The ith entry in the finger table of peer n contains the
      identity of the first peer that succeeds n by at least
      2^(numBitsInNodeId-i) on the Chord ring.  This peer is called the
      ith finger of peer n.  As an example, the first entry in the
      finger table of peer n contains a peer halfway around the Chord
      ring from peer n.  The purpose of the finger table is to
      accelerate lookups.

   n.id:  In this document, this abbreviation is used to refer to the
      Node-ID of peer n.

   O(g(n)):  Informally, saying that some equation f(n) = O(g(n)) means
      that f(n) is less than some constant multiple of g(n).  For the
      formal definition, please refer to [Weiss1998].

   Omega(g(n)):  Informally, saying that some equation f(n) =
      Omega(g(n)) means that f(n) is more than some constant multiple of
      g(n).  For the formal definition, please refer to [Weiss1998].

   Percentile:  The Pth (0<=P<=100) percentile of N values arranged in
      ascending order is obtained by first calculating the (ordinal)
      rank n=(P/100)*N, rounding the result to the nearest integer and
      then taking the value corresponding to that rank.

   Predecessor List:  A data structure containing the first r
      predecessors of a peer on the Chord ring.

   Successor List:  A data structure containing the first r successors
      of a peer on the Chord ring.

   Neighborhood Set:  A term used to refer to the set of peers included
      in the successor and predecessor lists of a given peer.







Maenpaa & Camarillo          Standards Track                    [Page 4]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


   Routing Table:  Contents of a given peer's routing table include the
      set of peers that the peer can use to route overlay messages.  The
      routing table is made up of the finger table, successor list, and
      predecessor list.

3.  Introduction to Stabilization in DHTs



   DHTs use stabilization routines to counter the undesirable effects of
   churn on routing.  The purpose of stabilization is to keep the
   routing information of each peer in the overlay consistent with the
   constantly changing overlay topology.  There are two alternative
   approaches to stabilization: periodic and reactive [Rhea2004].
   Periodic stabilization can either use a fixed stabilization rate or
   calculate the stabilization rate in an adaptive fashion.

3.1.  Reactive versus Periodic Stabilization



   In reactive stabilization, a peer reacts to the loss of a peer in its
   neighborhood set or to the appearance of a new peer that should be
   added to its neighborhood set by sending a copy of its neighbor table
   to all peers in the neighborhood set.  Periodic recovery, in
   contrast, takes place independently of changes in the neighborhood
   set.  In periodic recovery, a peer periodically shares its
   neighborhood set with each or a subset of the members of that set.

   The chord-reload algorithm [RFC6940] supports both reactive and
   periodic stabilization.  It has been shown in [Rhea2004] that
   reactive stabilization works well for small neighborhood sets (i.e.,
   small overlays) and moderate churn.  However, in large-scale (e.g.,
   1000 peers or more [Rhea2004]) or high-churn overlays, reactive
   stabilization runs the risk of creating a positive feedback cycle,
   which can eventually result in congestion collapse.  In [Rhea2004],
   it is shown that a 1000-peer overlay under churn uses significantly
   less bandwidth and has lower latencies when periodic stabilization is
   used than when reactive stabilization is used.  Although in the
   experiments carried out in [Rhea2004], reactive stabilization
   performed well when there was no churn, its bandwidth use was
   observed to jump dramatically under churn.  At higher churn rates and
   larger scale overlays, periodic stabilization uses less bandwidth and
   the resulting lower contention for the network leads to lower
   latencies.  For this reason, most DHTs, such as CAN [CAN], Chord
   [Chord], Pastry [Pastry], and Bamboo [Rhea2004], use periodic
   stabilization [Ghinita2006].  As an example, the first version of
   Bamboo used reactive stabilization, which caused Bamboo to suffer
   from degradation in performance under churn.  To fix this problem,
   Bamboo was modified to use periodic stabilization.





Maenpaa & Camarillo          Standards Track                    [Page 5]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


   In Chord, periodic stabilization is typically done both for
   successors and fingers.  An alternative strategy is analyzed in
   [Krishnamurthy2008].  In this strategy, called the "correction-on-
   change maintenance strategy", a peer periodically stabilizes its
   successors but does not do so for its fingers.  Instead, finger
   pointers are stabilized in a reactive fashion.  The results obtained
   in [Krishnamurthy2008] imply that although the correction-on-change
   strategy works well when churn is low, periodic stabilization
   outperforms the correction-on-change strategy when churn is high.

3.2.  Configuring Periodic Stabilization



   When periodic stabilization is used, one faces the problem of
   selecting an appropriate execution rate for the stabilization
   procedure.  If the execution rate of periodic stabilization is high,
   changes in the system can be quickly detected, but at the
   disadvantage of increased communication overhead.  Alternatively, if
   the stabilization rate is low and the churn rate is high, routing
   tables become inaccurate and DHT performance deteriorates.  Thus, the
   problem is setting the parameters so that the overlay achieves the
   desired reliability and performance even in challenging conditions,
   such as under heavy churn.  This naturally results in high cost
   during periods when the churn level is lower than expected, or
   alternatively, poor performance or even network partitioning in worse
   than expected conditions.

   In addition to selecting an appropriate stabilization interval,
   regardless of whether or not periodic stabilization is used, an
   appropriate size needs to be selected for the neighborhood set and
   for the finger table.

   The current approach is to configure overlays statically.  This works
   in situations where perfect information about the future is
   available.  In situations where the operating conditions of the
   network are known in advance and remain static throughout the
   lifetime of the system, it is possible to choose fixed optimal values
   for parameters such as stabilization rate, neighborhood set size and
   routing table size.  However, if the operating conditions (e.g., the
   size of the overlay and its churn rate) do not remain static but
   evolve with time, it is not possible to achieve both a low lookup
   failure rate and a low communication overhead by using fixed
   parameters [Ghinita2006].

   As an example, to configure the Chord DHT algorithm, one needs to
   select values for the following parameters: size of successor list,
   stabilization interval, and size of the finger table.  To select an
   appropriate value for the stabilization interval, one needs to know
   the expected churn rate and overlay size.  According to



Maenpaa & Camarillo          Standards Track                    [Page 6]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


   [Liben-Nowell2002], a Chord network in a ring-like state remains in a
   ring-like state as long as peers send Omega(square(log(N))) messages
   before N new peers join or N/2 peers fail.  Thus, in a 500-peer
   overlay churning at a rate such that one peer joins and one peer
   leaves the network every 30 seconds, an appropriate stabilization
   interval would be on the order of 93 s.  According to [Chord], the
   size of the successor list and finger table should be on the order of
   log(N).  Already a successor list of a modest size (e.g., log2(N) or
   2*log2(N), which is the successor list size used in [Chord]) makes it
   very unlikely that a peer will lose all of its successors, which
   would cause the Chord ring to become disconnected.  Thus, in a
   500-peer network each peer should maintain on the order of nine
   successors and fingers.  However, if the churn rate doubles and the
   network size remains unchanged, the stabilization rate should double
   as well.  That is, the appropriate maintenance interval would now be
   on the order of 46 s.  On the other hand, if the churn rate becomes,
   e.g., six-fold and the size of the network grows to 2000 peers, on
   the order of 11 fingers and successors should be maintained and the
   stabilization interval should be on the order of 42 s.  If one
   continued using the old values, this could result in inaccurate
   routing tables, network partitioning, and deteriorating performance.

3.3.  Adaptive Stabilization



   A self-tuning DHT takes into consideration the continuous evolution
   of network conditions and adapts to them.  In a self-tuning DHT, each
   peer collects statistical data about the network and dynamically
   adjusts its stabilization rate, neighborhood set size, and finger
   table size based on the analysis of the data [Ghinita2006].
   Reference [Mahajan2003] shows that by using self-tuning, it is
   possible to achieve high reliability and performance even in adverse
   conditions with low maintenance cost.  Adaptive stabilization has
   been shown to outperform periodic stabilization in terms of both
   lookup failures and communication overhead [Ghinita2006].

4.  Introduction to Chord



   Chord [Chord] is a structured P2P algorithm that uses consistent
   hashing to build a DHT out of several independent peers.  Consistent
   hashing assigns each peer and resource a fixed-length identifier.
   Peers use SHA-1 as the base hash function to generate the
   identifiers.  As specified in RELOAD base [RFC6940], the length of
   the identifiers is numBitsInNodeId=128 bits.  The identifiers are
   ordered on an identifier circle of size 2^numBitsInNodeId.  On the
   identifier circle, key k is assigned to the first peer whose
   identifier equals or follows the identifier of k in the identifier
   space.  The identifier circle is called the Chord ring.




Maenpaa & Camarillo          Standards Track                    [Page 7]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


   Different DHTs differ significantly in performance when bandwidth is
   limited.  It has been shown that when compared to other DHTs, the
   advantages of Chord include that it uses bandwidth efficiently and
   can achieve low lookup latencies at little cost [Li2004].

   A simple lookup mechanism could be implemented on a Chord ring by
   requiring each peer to only know how to contact its current successor
   on the identifier circle.  Queries for a given identifier could then
   be passed around the circle via the successor pointers until they
   encounter the first peer whose identifier is equal to or larger than
   the desired identifier.  Such a lookup scheme uses a number of
   messages that grows linearly with the number of peers.  To reduce the
   cost of lookups, Chord maintains also additional routing information;
   each peer n maintains a data structure with up to numBitsInNodeId
   entries, called the finger table.  The first entry in the finger
   table of peer n contains the peer halfway around the ring from peer
   n.  The second entry contains the peer that is 1/4th of the way
   around, the third entry the peer that is 1/8th of the way around,
   etc.  In other words, the ith entry in the finger table at peer n
   contains the identity of the first peer s that succeeds n by at least
   2^(numBitsInNodeId-i) on the Chord ring.  This peer is called the ith
   finger of peer n.  The interval between two consecutive fingers is
   called a finger interval.  The ith finger interval of peer n covers
   the range [n.id + 2^(numBitsInNodeId-i), n.id + 2^(numBitsInNodeId-
   i+1)) on the Chord ring.  In an N-peer network, each peer maintains
   information about O(log(N)) other peers in its finger table.  As an
   example, if N=100000, it is sufficient to maintain 17 fingers.

   Chord needs all peers' successor pointers to be up to date in order
   to ensure that lookups produce correct results as the set of
   participating peers changes.  To achieve this, peers run a
   stabilization protocol periodically in the background.  The
   stabilization protocol of the original Chord algorithm uses two
   operations: successor stabilization and finger stabilization.
   However, the Chord algorithm of RELOAD base defines two additional
   stabilization components, as will be discussed below.

   To increase robustness in the event of peer failures, each Chord peer
   maintains a successor list of size r, containing the peer's first r
   successors.  The benefit of successor lists is that if each peer
   fails independently with probability p, the probability that all r
   successors fail simultaneously is only p^r.

   The original Chord algorithm maintains only a single predecessor
   pointer.  However, multiple predecessor pointers (i.e., a predecessor
   list) can be maintained to speed up recovery from predecessor
   failures.  The routing table of a peer consists of the successor
   list, finger table, and predecessor list.



Maenpaa & Camarillo          Standards Track                    [Page 8]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


5.  Extending Chord-Reload to Support Self-Tuning



   This section describes how the mandatory-to-implement chord-reload
   algorithm defined in RELOAD base [RFC6940] can be extended to support
   self-tuning.

   The chord-reload algorithm supports both reactive and periodic
   recovery strategies.  When the self-tuning mechanisms defined in this
   document are used, the periodic recovery strategy is used.  Further,
   chord-reload specifies that at least three predecessors and three
   successors need to be maintained.  When the self-tuning mechanisms
   are used, the appropriate sizes of the successor list and predecessor
   list are determined in an adaptive fashion based on the estimated
   network size, as will be described in Section 6.

   As specified in RELOAD base [RFC6940], each peer maintains a
   stabilization timer.  When the stabilization timer fires, the peer
   restarts the timer and carries out the overlay stabilization routine.
   Overlay stabilization has four components in chord-reload:

   1.  Update the neighbor table.  We refer to this as "neighbor
       stabilization".

   2.  Refreshing the finger table.  We refer to this as "finger
       stabilization".

   3.  Adjusting finger table size.

   4.  Detecting partitioning.  We refer to this as "strong
       stabilization".

   As specified in RELOAD base [RFC6940], a peer sends periodic messages
   as part of the neighbor stabilization, finger stabilization, and
   strong stabilization routines.  In neighbor stabilization, a peer
   periodically sends an Update request to every peer in its connection
   table.  The default time is every ten minutes.  In finger
   stabilization, a peer periodically searches for new peers to include
   in its finger table.  This time defaults to one hour.  This document
   specifies how the neighbor stabilization and finger stabilization
   intervals can be determined in an adaptive fashion based on the
   operating conditions of the overlay.  The subsections below describe
   how this document extends the four components of stabilization.

5.1.  Update Requests



   As described in RELOAD base [RFC6940], the neighbor and finger
   stabilization procedures are implemented using Update requests.
   RELOAD base defines three types of Update requests: 'peer_ready',



Maenpaa & Camarillo          Standards Track                    [Page 9]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


   'neighbors', and 'full'.  Regardless of the type, all Update requests
   include an 'uptime' field.  The self-tuning extensions require
   information on the uptimes of peers in the routing table.  The sender
   of an Update request includes its current uptime (in seconds) in the
   'uptime' field.  Regardless of the type, all Update requests MUST
   include an 'uptime' field.

   When self-tuning is used, each peer decides independently the
   appropriate size for the successor list, predecessor list, and finger
   table.  Thus, the 'predecessors', 'successors', and 'fingers' fields
   included in RELOAD Update requests are of variable length.  As
   specified in RELOAD [RFC6940], variable-length fields are on the wire
   preceded by length bytes.  In the case of the successor list,
   predecessor list, and finger table, there are two length bytes
   (allowing lengths up to 2^16-1).  The number of NodeId structures
   included in each field can be calculated based on the length bytes
   since the size of a single NodeId structure is 16 bytes.  If a peer
   receives more entries than fit into its successor list, predecessor
   list, or finger table, the peer MUST ignore the extra entries.  A
   peer may also receive less entries than it currently has in its own
   data structure.  In that case, it uses the received entries to update
   only a subset of the entries in its data structure.  As an example, a
   peer that has a successor list of size 8 may receive a successor list
   of size 4 from its immediate successor.  In that case, the received
   successor list can only be used to update the first few successors on
   the peer's successor list.  The rest of the successors will remain
   intact.

5.2.  Neighbor Stabilization



   In the neighbor stabilization operation of chord-reload, a peer
   periodically sends an Update request to every peer in its connection
   table.  In a small, low-churn overlay, the amount of traffic this
   process generates is typically acceptable.  However, in a large-scale
   overlay churning at a moderate or high churn rate, the traffic load
   may no longer be acceptable since the size of the connection table is
   large and the stabilization interval relatively short.  The self-
   tuning mechanisms described in this document are especially designed
   for overlays of the latter type.  Therefore, when the self-tuning
   mechanisms are used, each peer only sends a periodic Update request
   to its first predecessor and first successor on the Chord ring; it
   MUST NOT send Update requests to others.

   The neighbor stabilization routine is executed when the stabilization
   timer fires.  To begin the neighbor stabilization routine, a peer
   sends an Update request to its first successor and its first
   predecessor.  The type of the Update request MUST be 'neighbors'.
   The Update request includes the successor and predecessor lists of



Maenpaa & Camarillo          Standards Track                   [Page 10]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


   the sender.  If a peer receiving such an Update request learns from
   the predecessor and successor lists included in the request that new
   peers can be included in its neighborhood set, it sends Attach
   requests to the new peers.

   After a new peer has been added to the predecessor or successor list,
   an Update request of type 'peer_ready' is sent to the new peer.  This
   allows the new peer to insert the sender into its neighborhood set.

5.3.  Finger Stabilization



   Chord-reload specifies two alternative methods for searching for new
   peers to the finger table.  Both of the alternatives can be used with
   the self-tuning extensions defined in this document.

   Immediately after a new peer has been added to the finger table, a
   Probe request is sent to the new peer to fetch its uptime.  The
   'requested_info' field of the Probe request MUST be set to contain
   the ProbeInformationType 'uptime' defined in RELOAD base [RFC6940].

5.4.  Adjusting Finger Table Size



   The chord-reload algorithm defines how a peer can make sure that the
   finger table is appropriately sized to allow for efficient routing.
   Since the self-tuning mechanisms specified in this document produce a
   network size estimate, this estimate can be directly used to
   calculate the optimal size for the finger table.  This mechanism is
   used instead of the one specified by chord-reload.  A peer uses the
   network size estimate to determine whether it needs to adjust the
   size of its finger table each time when the stabilization timer
   fires.  The way this is done is explained in Section 6.2.

5.5.  Detecting Partitioning



   This document does not require any changes to the mechanism chord-
   reload uses to detect network partitioning.

5.6.  Leaving the Overlay



   As specified in RELOAD base [RFC6940], a leaving peer SHOULD send a
   Leave request to all members of its neighbor table prior to leaving
   the overlay.  The 'overlay_specific_data' field MUST contain the
   ChordLeaveData structure.  The Leave requests that are sent to
   successors contain the predecessor list of the leaving peer.  The
   Leave requests that are sent to the predecessors contain the
   successor list of the leaving peer.  If a given successor can
   identify better predecessors (that is, predecessors that are closer
   to it on the Chord ring than its existing predecessors) than are



Maenpaa & Camarillo          Standards Track                   [Page 11]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


   already included in its predecessor lists by investigating the
   predecessor list it receives from the leaving peer, it sends Attach
   requests to them.  Similarly, if a given predecessor identifies
   better successors by investigating the successor list it receives
   from the leaving peer, it sends Attach requests to them.

6.  Self-Tuning Chord Parameters



   This section specifies how to determine an appropriate stabilization
   rate and routing table size in an adaptive fashion.  The proposed
   mechanism is based on [Mahajan2003], [Liben-Nowell2002], and
   [Ghinita2006].  To calculate an appropriate stabilization rate, the
   values of three parameters must be estimated: overlay size N, failure
   rate U, and join rate L.  To calculate an appropriate routing table
   size, the estimated network size N can be used.  Peers in the overlay
   MUST recalculate the values of the parameters to self-tune the chord-
   reload algorithm at the end of each stabilization period before
   restarting the stabilization timer.

6.1.  Estimating Overlay Size



   Techniques for estimating the size of an overlay network have been
   proposed, for instance, in [Mahajan2003], [Horowitz2003],
   [Kostoulas2005], [Binzenhofer2006], and [Ghinita2006].  In Chord, the
   density of peer identifiers in the neighborhood set can be used to
   produce an estimate of the size of the overlay, N [Mahajan2003].
   Since peer identifiers are picked randomly with uniform probability
   from the numBitsInNodeId-bit identifier space, the average distance
   between peer identifiers in the successor set is
   (2^numBitsInNodeId)/N.

   To estimate the overlay network size, a peer computes the average
   inter-peer distance d between the successive peers starting from the
   most distant predecessor and ending to the most distant successor in
   the successor list.  The estimated network size is calculated as:

                         2^numBitsInNodeId
                    N = -------------------
                                d

   This estimate has been found to be accurate within 15% of the real
   network size [Ghinita2006].  Of course, the size of the neighborhood
   set affects the accuracy of the estimate.








Maenpaa & Camarillo          Standards Track                   [Page 12]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


   During the join process, a joining peer fills its routing table by
   sending a series of Ping and Attach requests, as specified in RELOAD
   base [RFC6940].  Thus, a joining peer immediately has enough
   information at its disposal to calculate an estimate of the network
   size.

6.2.  Determining Routing Table Size



   As specified in RELOAD base [RFC6940], the finger table must contain
   at least 16 entries.  When the self-tuning mechanisms are used, the
   size of the finger table MUST be set to max(ceiling(log2(N)), 16)
   using the estimated network size N.

   The size of the successor list MUST be set to a maximum of
   ceiling(log2(N)).  An implementation can place a lower limit on the
   size of the successor list.  As an example, the implementation might
   require the size of the successor list to be always at least three.

   The size of the predecessor list MUST be set to ceiling(log2(N)).

6.3.  Estimating Failure Rate



   A typical approach is to assume that peers join the overlay according
   to a Poisson process with rate L and leave according to a Poisson
   process with rate parameter U [Mahajan2003].  The value of U can be
   estimated using peer failures in the finger table and neighborhood
   set [Mahajan2003].  If peers fail with rate U, a peer with M unique
   peer identifiers in its routing table should observe K failures in
   time K/(M*U).  Every peer in the overlay maintains a history of the
   last K failures.  The current time is inserted into the history when
   the peer joins the overlay.  The estimate of U is calculated as:

                             k
                     U = --------,
                          M * Tk

   where M is the number of unique peer identifiers in the routing
   table, Tk is the time between the first and the last failure in the
   history, and k is the number of failures in the history.  If k is
   smaller than K, the estimate is computed as if there was a failure at
   the current time.  It has been shown that an estimate calculated in a
   similar manner is accurate within 17% of the real value of U
   [Ghinita2006].

   The size of the failure history K affects the accuracy of the
   estimate of U.  One can increase the accuracy by increasing K.
   However, this has the side effect of decreasing responsiveness to
   changes in the failure rate.  On the other hand, a small history size



Maenpaa & Camarillo          Standards Track                   [Page 13]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


   may cause a peer to overreact each time a new failure occurs.  In
   [Ghinita2006], K is set to 25% of the routing table size.  Use of
   this value is RECOMMENDED.

6.3.1.  Detecting Failures



   A new failure is inserted to the failure history in the following
   cases:

   1.  A Leave request is received from a neighbor.

   2.  A peer fails to reply to a Ping request sent in the situation
       explained below.  If no packets have been received on a
       connection during the past 2*Tr seconds (where Tr is the
       inactivity timer defined by Interactive Connectivity
       Establishment (ICE) [RFC5245]), a RELOAD Ping request MUST be
       sent to the remote peer.  RELOAD mandates the use of Session
       Traversal Utilities for NAT (STUN) [RFC5389] for keepalives.
       STUN keepalives take the form of STUN Binding Indication
       transactions.  As specified in ICE [RFC5245], a peer sends a STUN
       Binding Indication if there has been no packet sent on a
       connection for Tr seconds.  Tr is configurable and has a default
       of 15 seconds.  Although STUN Binding Indications do not generate
       a response, the fact that a peer has failed can be learned from
       the lack of packets (Binding Indications or application protocol
       packets) received from the peer.  If the remote peer fails to
       reply to the Ping request, the sender should consider the remote
       peer to have failed.

   As an alternative to relying on STUN keepalives to detect peer
   failure, a peer could send additional, frequent RELOAD messages to
   every peer in its connection table.  These messages could be Update
   requests, in which case they would serve two purposes: detecting peer
   failure and stabilization.  However, as the cost of this approach can
   be very high in terms of bandwidth consumption and traffic load,
   especially in large-scale overlays experiencing churn, its use is NOT
   RECOMMENDED
.

6.4.  Estimating Join Rate



   Reference [Ghinita2006] proposes that a peer can estimate the join
   rate based on the uptime of the peers in its routing table.  An
   increase in peer join rate will be reflected by a decrease in the
   average age of peers in the routing table.  Thus, each peer
   maintained an array of the ages of the peers in its routing table
   sorted in increasing order.  Using this information, an estimate of
   the global peer join rate L is calculated as:




Maenpaa & Camarillo          Standards Track                   [Page 14]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


                                  N
                    L = ----------------------,
                         Ages[floor(rsize/2)]

   where Ages is an array containing the ages of the peers in the
   routing table sorted in increasing order and rsize is the size of the
   routing table.  It has been shown that the estimate obtained by using
   this method is accurate within 22% of the real join rate
   [Ghinita2006].  Of course, the size of the routing table affects the
   accuracy.

   In order for this mechanism to work, peers need to exchange
   information about the time they have been present in the overlay.
   Peers receive the uptimes of their successors and predecessors during
   the stabilization operations since all Update requests carry uptime
   values.  A joining peer learns the uptime of the admitting peer since
   it receives an Update from the admitting peer during the join
   procedure.  Peers learn the uptimes of new fingers since they can
   fetch the uptime using a Probe request after having attached to the
   new finger.

6.5.  Estimate Sharing



   To improve the accuracy of network size, join rate, and leave rate
   estimates, peers share their estimates.  When the stabilization timer
   fires, a peer selects number-of-peers-to-probe random peers from its
   finger table and send each of them a Probe request.  The targets of
   Probe requests are selected from the finger table rather than from
   the neighbor table since neighbors are likely to make similar errors
   when calculating their estimates.  The number-of-peers-to-probe is a
   new element in the overlay configuration document.  It is defined in
   Section 7.  Both the Probe request and the answer returned by the
   target peer MUST contain a new message extension whose
   MessageExtensionType is 'self_tuning_data'.  This extension type is
   defined in Section 9.1.  The 'extension_contents' field of the
   MessageExtension structure MUST contain a SelfTuningData structure:

               struct {
                 uint32                   network_size;
                 uint32                   join_rate;
                 uint32                   leave_rate;
               } SelfTuningData;









Maenpaa & Camarillo          Standards Track                   [Page 15]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


   The contents of the SelfTuningData structure are as follows:

   network_size

      The latest network size estimate calculated by the sender.

   join_rate

      The latest join rate estimate calculated by the sender.

   leave_rate

      The latest leave rate estimate calculated by the sender.

   The join and leave rates are expressed as joins or failures per 24
   hours.  As an example, if the global join rate estimate a peer has
   calculated is 0.123 peers/s, it would include in the 'join_rate'
   field the ceiling of the value 10627.2 (24*60*60*0.123 = 10627.2),
   that is, the value 10628.

   The 'type' field of the MessageExtension structure MUST be set to
   contain the value 'self_tuning_data'.  The 'critical' field of the
   structure MUST be set to False.

   A peer stores all estimates it receives in Probe requests and answers
   during a stabilization interval.  When the stabilization timer fires,
   the peer calculates the estimates to be used during the next
   stabilization interval by taking the 75th percentile (i.e., third
   quartile) of a data set containing its own estimate and the received
   estimates.

   The default value for number-of-peers-to-probe is 4.  This default
   value is recommended to allow a peer to receive a sufficiently large
   set of estimates from other peers.  With a value of 4, a peer
   receives four estimates in Probe answers.  On the average, each peer
   also receives four Probe requests each carrying an estimate.  Thus,
   on the average, each peer has nine estimates (including its own) that
   it can use at the end of the stabilization interval.  A value smaller
   than 4 is NOT RECOMMENDED to keep the number of received estimates
   high enough.  As an example, if the value were 2, there would be
   peers in the overlay that would only receive two estimates during a
   stabilization interval.  Such peers would only have three estimates
   available at the end of the interval, which may not be reliable
   enough since even a single exceptionally high or low estimate can
   have a large impact.






Maenpaa & Camarillo          Standards Track                   [Page 16]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


6.6.  Calculating the Stabilization Interval



   According to [Liben-Nowell2002], a Chord network in a ring-like state
   remains in a ring-like state as long as peers send
   Omega(square(log(N))) messages before N new peers join or N/2 peers
   fail.  We can use the estimate of peer failure rate, U, to calculate
   the time Tf in which N/2 peers fail:

                                  1
                           Tf = ------
                                 2*U

   Based on this estimate, a stabilization interval Tstab-1 is
   calculated as:

                                           Tf
                           Tstab-1 = -----------------
                                      square(log2(N))

   On the other hand, the estimated join rate L can be used to calculate
   the time in which N new peers join the overlay.  Based on the
   estimate of L, a stabilization interval Tstab-2 is calculated as:

                                               N
                            Tstab-2 = ---------------------
                                       L * square(log2(N))

   Finally, the actual stabilization interval Tstab that is used can be
   obtained by taking the minimum of Tstab-1 and Tstab-2.

   The results obtained in [Maenpaa2009] indicate that making the
   stabilization interval too small has the effect of making the overlay
   less stable (e.g., in terms of detected loops and path failures).
   Thus, a lower limit should be used for the stabilization period.
   Based on the results in [Maenpaa2009], a lower limit of 15 s is
   RECOMMENDED, since using a stabilization period smaller than this
   will with a high probability cause too much traffic in the overlay.

7.  Overlay Configuration Document Extension



   This document extends the RELOAD overlay configuration document by
   adding one new element, "number-of-peers-to-probe", inside each
   "configuration" element.

   self-tuning:number-of-peers-to-probe:  The number of fingers to which
      Probe requests are sent to obtain their network size, join rate,
      and leave rate estimates.  The default value is 4.




Maenpaa & Camarillo          Standards Track                   [Page 17]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


   The RELAX NG grammar for this element is:

   namespace self-tuning = "urn:ietf:params:xml:ns:p2p:self-tuning"

   parameter &= element self-tuning:number-of-peers-to-probe {
   xsd:unsignedInt }?

   This namespace is added into the <mandatory-extension> element in the
   overlay configuration file.

8.  Security Considerations



   In the same way as malicious or compromised peers implementing the
   RELOAD base protocol [RFC6940] can advertise false network metrics or
   distribute false routing table information for instance in RELOAD
   Update messages, malicious peers implementing this specification may
   share false join rate, leave rate, and network size estimates.  For
   such attacks, the same security concerns apply as in the RELOAD base
   specification.  In addition, as long as the amount of malicious peers
   in the overlay remains modest, the statistical mechanisms applied in
   Section 6.5 (i.e., the use of 75th percentiles) to process the shared
   estimates a peer obtains help ensure that estimates that are clearly
   different from (i.e., larger or smaller than) other received
   estimates will not significantly influence the process of adapting
   the stabilization interval and routing table size.  However, it
   should be noted that if an attacker is able to impersonate a high
   number of other peers in the overlay in strategic locations, it may
   be able to send a high enough number of false estimates to a victim
   and therefore influence the victim's choice of a stabilization
   interval.

9.  IANA Considerations



9.1.  Message Extensions



   This document introduces one additional extension to the "RELOAD
   Extensions Registry":

                  +------------------+-------+---------------+
                  | Extension Name   |  Code | Specification |
                  +------------------+-------+---------------+
                  | self_tuning_data |   0x3 |      RFC 7363 |
                  +------------------+-------+---------------+


   The contents of the extension are defined in Section 6.5.





Maenpaa & Camarillo          Standards Track                   [Page 18]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


9.2.  New Overlay Algorithm Type



   This document introduces one additional overlay algorithm type to the
   "RELOAD Overlay Algorithm Types" registry:

                  +-------------------+-----------+
                  | Algorithm Name    | Reference |
                  +-------------------+-----------+
                  | CHORD-SELF-TUNING | RFC 7363  |
                  +-------------------+-----------+


9.3.  A New IETF XML Registry



   This document registers one new URI for the self-tuning namespace in
   the "ns" subregistry of the IETF XML registry defined in [RFC3688].

   URI: urn:ietf:params:xml:ns:p2p:self-tuning

   Registrant Contact: The IESG

   XML: N/A, the requested URI is an XML namespace

10.  Acknowledgments



   The authors would like to thank Jani Hautakorpi for his contributions
   to the document.  The authors would also like to thank Carlos
   Bernardos, Martin Durst, Alissa Cooper, Tobias Gondrom, and Barry
   Leiba for their comments on the document.

11.  References



11.1.  Normative References



   [RFC2119]  Bradner, S., "Key words for use in RFCs to Indicate
              Requirement Levels", BCP 14, RFC 2119, March 1997.

   [RFC5245]  Rosenberg, J., "Interactive Connectivity Establishment
              (ICE): A Protocol for Network Address Translator (NAT)
              Traversal for Offer/Answer Protocols", RFC 5245, April
              2010.

   [RFC5389]  Rosenberg, J., Mahy, R., Matthews, P., and D. Wing,
              "Session Traversal Utilities for NAT (STUN)", RFC 5389,
              October 2008.






Maenpaa & Camarillo          Standards Track                   [Page 19]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


   [RFC6940]  Jennings, C., Lowekamp, B., Rescorla, E., Baset, S., and
              H. Schulzrinne, "REsource LOcation And Discovery (RELOAD)
              Base Protocol", RFC 6940, January 2014.

11.2.  Informative References



   [Binzenhofer2006]
              Binzenhofer, A., Kunzmann, G., and R. Henjes, "A Scalable
              Algorithm to Monitor Chord-Based P2P Systems at Runtime",
              In Proceedings of the 20th IEEE International Parallel and
              Distributed Processing Symposium (IPDPS), pp. 1-8, April
              2006.

   [CAN]      Ratnasamy, S., Francis, P., Handley, M., Karp, R., and S.
              Schenker, "A Scalable Content-Addressable Network", In
              Proceedings of the 2001 Conference on Applications,
              Technologies, Architectures and Protocols for Computer
              Communications, pp. 161-172, August 2001.

   [Chord]    Stoica, I., Morris, R., Liben-Nowell, D., Karger, D.,
              Kaashoek, M., Dabek, F., and H. Balakrishnan, "Chord: A
              Scalable Peer-to-peer Lookup Service for Internet
              Applications", IEEE/ACM Transactions on Networking, Volume
              11, Issue 1, pp. 17-32, February 2003.

   [Ghinita2006]
              Ghinita, G. and Y. Teo, "An Adaptive Stabilization
              Framework for Distributed Hash Tables", In Proceedings of
              the 20th IEEE International Parallel and Distributed
              Processing Symposium (IPDPS), pp. 29-38, April 2006.

   [Horowitz2003]
              Horowitz, K. and D. Malkhi, "Estimating Network Size from
              Local Information", Information Processing Letters, Volume
              88, Issue 5, pp. 237-243, December 2003.

   [Kostoulas2005]
              Kostoulas, D., Psaltoulis, D., Gupta, I., Birman, K., and
              A. Demers, "Decentralized Schemes for Size Estimation in
              Large and Dynamic Groups", In Proceedings of the 4th IEEE
              International Symposium on Network Computing and
              Applications, pp. 41-48, July 2005.









Maenpaa & Camarillo          Standards Track                   [Page 20]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


   [Krishnamurthy2008]
              Krishnamurthy, S., El-Ansary, S., Aurell, E., and S.
              Haridi, "Comparing Maintenance Strategies for Overlays",
              In Proceedings of the 16th Euromicro Conference on
              Parallel, Distributed and Network-Based Processing, pp.
              473-482, February 2008.

   [Li2004]   Li, J., Strinbling, J., Gil, T., Morris, R., and M.
              Kaashoek, "Comparing the Performance of Distributed Hash
              Tables Under Churn", Peer-to-Peer Systems III, Volume 3279
              of Lecture Notes in Computer Science, Springer, pp. 87-99,
              February 2005.

   [Liben-Nowell2002]
              Liben-Nowell, D., Balakrishnan, H., and D. Karger,
              "Observations on the Dynamic Evolution of Peer-to-Peer
              Networks", In Proceedings of the 1st International
              Workshop on Peer-to-Peer Systems (IPTPS), pp. 22-33, March
              2002.

   [Maenpaa2009]
              Maenpaa, J. and G. Camarillo, "A Study on Maintenance
              Operations in a Chord-Based Peer-to-Peer Session
              Initiation Protocol Overlay Network", In Proceedings of
              the 23rd IEEE International Parallel and Distributed
              Processing Symposium (IPDPS), pp. 1-9, May 2009.

   [Mahajan2003]
              Mahajan, R., Castro, M., and A. Rowstron, "Controlling the
              Cost of Reliability in Peer-to-Peer Overlays", In
              Proceedings of the 2nd International Workshop on Peer-to-
              Peer Systems (IPTPS), pp. 21-32, February 2003.

   [Pastry]   Rowstron, A. and P. Druschel, "Pastry: Scalable,
              Decentralized Object Location and Routing for Large-Scale
              Peer-to-Peer Systems", In Proceedings of the IFIP/ACM
              International Conference on Distributed Systems Platforms,
              pp. 329-350, November 2001.

   [RFC3688]  Mealling, M., "The IETF XML Registry", BCP 81, RFC 3688,
              January 2004.

   [Rhea2004]
              Rhea, S., Geels, D., Roscoe, T., and J. Kubiatowicz,
              "Handling Churn in a DHT", In Proceedings of the USENIX
              Annual Technical Conference, pp. 127-140, June 2004.





Maenpaa & Camarillo          Standards Track                   [Page 21]

RFC 7363               Self-Tuning DHT for RELOAD         September 2014


   [Weiss1998]
              Weiss, M., "Data Structures and Algorithm Analysis in
              C++", Addison-Wesley Longman Publishing Co., Inc., 2nd
              Edition, ISBN 0201361221, 1998.

Authors' Addresses



   Jouni Maenpaa
   Ericsson
   Hirsalantie 11
   Jorvas  02420
   Finland

   EMail: Jouni.Maenpaa@ericsson.com


   Gonzalo Camarillo
   Ericsson
   Hirsalantie 11
   Jorvas  02420
   Finland

   EMail: Gonzalo.Camarillo@ericsson.com




























Maenpaa & Camarillo          Standards Track                   [Page 22]