20120412 searching techniques in peer to peer networks

Author: Xiuqi Li and Jie Wu
Presenter: Zia Ush Shamszaman
ANLAB, ICE, HUFS

 Survey of major searching techniques in peer-to-peer (P2P)
networks.
 Concept of P2P networks and the methods for classifying
different P2P networks .
 Various searching techniques in unstructured P2P systems, strictly
structured P2P systems, and loosely structured P2P systems has
been discussed.
 Searching in unstructured P2Ps covers both blind search schemes
and informed search scheme.
 Searching in Strictly structured P2Ps focuses on hierarchical
Distributed Hash Table (DHT) P2Ps and
 Non-DHT P2Ps and non-hierarchical DHT P2Ps is briefly
overviewed.

2

 P2P networks are overlay networks on top of Internet, where
nodes are end systems in the Internet and maintain information
about a set of other nodes (called neighbors) in the P2P.

 P2P networks offer the following benefits
 They do not require any special administration or financial
arrangements.
 They are self-organized and adaptive. Peers may come and go freely.
P2P systems handle these events automatically.
 They can gather and harness the tremendous computation and storage
resources on computers across the Internet.
 They are distributed and decentralized. Therefore, they are
potentially fault-tolerant and load-balanced.

3

 P2P networks can be classified based on the control over
data location and network topology.
 There are three categories:
 Unstructured: In an unstructured P2P network such as
Gnutella, no rule exists which defines where data is stored and
the network topology is arbitrary.

 Loosely structured: In a loosely structured network such as
Freenet and Symphony, the overlay structure and the data
location are not precisely determined.

 Highly structured: In a highly structured P2P network such as
Chord, both the network architecture and the data placement
are precisely specified.

4

 P2P networks can also be classified as centralized and decentralized
 In a centralized P2P such as Napster, a central directory of object location, ID
assignment, etc. is maintained in a single location.
 Decentralized P2Ps adopt a distributed directory structure.
These systems can be further divided into
 Purely decentralized systems, such as Gnutella and Chord, peers are totally equal.
 Hybrid systems, some peers called dominating nodes or super-peers serve the search
request of other regular peers.

 Another classification of P2P systems is hierarchical & non-hierarchical
based on whether the overlay structure is a hierarchy or not.
 All hybrid systems and few purely decentralized systems such as Kelips, are
hierarchical systems. Hierarchical systems provide good scalability,
opportunity to take advantage of node heterogeneity, and high routing
efficiency
 Most purely decentralized systems have flat overlays and are non-hierarchical
systems. Non-hierarchical systems offer load-balance and highresilience

5

 Searching means locating desired data.
 Most existing P2P systems support thesimple object lookup by key
or identifier.
 Some existing P2P systems can handle more complex keyword
queries, which find documents containing keywords in queries.
 More than one copy of an object may exist in a P2P system.
There may be more than one document that contains desired
keywords.
 Some P2P systems are interested in a single data item; others are
interested in all data items or as many data items as possible
that satisfy a given condition.
 Most searching techniques are forwarding-based. Starting with
the requesting node, a query is forwarded (or routed) to the
desired node/s.

6

 High-quality query results
 Minimal routing state maintained per node
 High routing efficiency
 Load balance
 Resilience to node failures
 Support of complex queries

7

 The quality of query results is application dependent.
 Generally, it is measured by the number of results and
relevance.

 The routing state refers to the number of neighbors each
node maintains.
 The routing efficiency is generally measured by the
number of overlay hops per query.
 In some systems, it is also evaluated using the number of
messages per query.
 Different searching techniques make different trade-offs
between these desired characteristics.

8

Yang and Garcia-Molina borrowed the idea of
iterative deepening from artificial intelligence.

 The querying node periodically issues a sequence of BFS
searches with increasing depth limits D1 < D2 < … < Di.

 The query is terminated when the query result is satisfied
or when the maximum depth limit D has been reached.

 All nodes use the same sequence of depth limits called
Policy: set of depths {0,2,4,5} and,

11

 Iterative Deepening – {0, 2, 4, 5}, 3

Holding Frozen Query Already Processed Query Unaware of Query

Standard Random Walker
 Forward the query to a randomly chosen neighbor at each
step Each message a walker.
 Cut message overhead
 Increase query searching delay(#hops)
k-walkers
 The requesting node sends k query messages and each
query message takes its own random walk
 Periodically, when a node receives a query, it checks with
source node to see if query has been satisfied
 k walkers after T steps should reach roughly the same
number of nodes as 1 walker after kT steps
 So cut delay by a factor of k.
 To decrease delay, increase walkers

16

Why
shouldn’t I
find a song? A sends a walker
to find song.mp3
that is stored on B

 TTL-based or Hop Count
 Checking: the walker periodically checks with the original
requestor before walking to the next node (again use a
large TTL, just to prevent loops)

Experiments show that
checking once at every 4th step strikes a good balance
between the overhead of the checking message and the
benefits of checking

18

 Directed Breadth First Search
 Source Only sends queries to good neighbors
 Good neighbors might have
 Produced results in the past
 Low latency
 Lowest hop count for results
 They have good neighbors
 Highest traffic neighbors
 They’re stable
 Shortest message queue
 Routed as normal BFS after first hop

Efficient Search - Methods


 Idea is that a node maintains information about
what files neighbors, and possibly neighbor’s
neighbors store. “radius of knowledge”
 All nodes know about radius, and know about
policy
 Policy lists which levels will respond and which will
ignore messages
 Servents look at TTL/Hops to determine if they
process or ignore.
 Memory issue of maintaining lists
 Size is far below a megabyte for radius < 5
 Network issue of building list
 Hit from extra packets

Nodes at level 1
Each node has a radius of
respond with
2 – knows about the files
information about
on its neighbors and
levels 1, 2 and 3, … neighbor’s neighbors
and forward to next
level Policy is that levels 1, 4
… and 7 respond

…
…
Searches move to
… levels 2 and 3, which
ignore and forward reaches
When search
level 4, it responds with
information about Layers 5 level67
Finally,
and …
levels 4, 5, and 6, simply ignore and own
then
responds with its
forward.
forwards the messages. and terminates
data,
the query.

 Joining a new node: sends a join message with TTL=r and
all the nodes within r hops update their indices.
 Join message contains the metadata about the joining
node.
 When a node receives this join message it, in turn, send
join message containing its meta data directly to the new
node. New node updates its indices.
 Node dies: Other nodes update their indices based on the
timeouts.
 Updating the node: When a node updates its collection, his
node will send out a small update message with TTL= r,
containing the metadata of the affected item. All nodes
receiving this message subsequently update their index.

 The objective of a Routing Index (RI) is to allow
a node to select the “best” neighbors to send a
query.

 A RI is a data structure that, given a query, retur
ns a list of neighbors, ranked according to their
goodness for the query.

 Each node has a local index for quickly finding lo
cal documents when a query is received. Nodes
also have a CRI containing
 the number of documents along each path
 the number of documents on each topic

28

 For A, there are
 100 documents available from B (and its descendents)
 20 belong to Database category

 10 belong to Theory category

 30 belong to Languages category

 “Goodness” of a neighbor
CRI ( s i )
Number Of Documents
i Number Of Documents

CRI(si) is the value for the cell at the column for topic si and at the
row for a neighbor

 For documents of “databases” and “languages”
20 30
Goodness ( B ) 100 6
100 100
0 50
Goodness ( C ) 1000 0
1000 1000
100 150
Goodness ( D ) 200 75
200 200

New connection

RI propagation D+A+J

D+A+I

 Attenuated Bloom Filters are extensions to bloom filters.
 Bloom filters are often used to approximately and
efficiently summarize elements in a set.
 Assumes that each stored document has many replicas
spread over the P2P network.
 Documents are queried by names.
 It intends to quickly find replicas close to the query source
with high probability.
 This is achieved by approximately
 Summarizing the documents that likely exist in nearby
nodes.
 However, the approach alone fails to find replicas far away
from the query source.

32

 In a strictly structured system, the neighbor relationship
between peers and data locations is strictly defined.
 Searching in such systems is therefore determined by the
particular network architecture.
 Among the strictly structured systems, some implement a
distributed hash table (DHT) using different data
structures. Others do not provide a DHT interface.
 Some DHT P2Psystems have flat overlay structures; others
have hierarchical overlay structures.
 A DHT is a hash table whose table entries are distributed among
different p eers located in arbitrary locations.
 Each data item is hashed to a unique numeric key. Each node is
also hashed to a unique ID in the same key space

33

 Different non-hierarchical DHT P2Ps use different flat data
structures to implement the DHT.
 These flat data structures include ring, mesh, hypercube, and
other special graphs such as de Bruijn graph.
 Chord uses a ring data structure
 Node IDs form a ring.
 Each node keeps a finger table that contains the IP addresses of nodes

 Pastry uses a tree-based data structure which can be
considered as a generalization of a hypercube.
 To shorten the routing latency, each pastry node also keeps a routing
table of pointers to other nodes in the ID space
 Some other examples are Koorde, Viceroy, and Cycloid etc.

34

 All hierarchical DHT P2Ps organize peers into different
groups or clusters.
 Each group forms its own overlay.
 All groups together form the entire hierarchical overlay.
 Typically the overlay hierarchies are two-tier or three-tier.
 They differ mainly in the number of groups in each tier,
the overlay structure formed by each group.
 Superpeers/dominating nodes generally contribute more
 computing resources, are more stable, and take more
responsibility in routing than regular peers.

 Example of this category is Kelips and Coral.

35

 Kelips is composed of k virtual affinity groups with group
IDs.
 IP address and port number of a node n is hashed to a
group ID of the group to which the node n belongs.
 The consistent hashing function SHA-1 provides a good
balance of group members with high probability.
 Each file name is mapped to a group using the same SHA-1
function.
 Inside a group, a file is stored in a randomly chosen group
member, called the file’s homenode.
 Kelips offers load balance in the same group and among
different groups

36

Each node n in an affinity group “g” keeps in the memory
the following routing state:
 View of the belonging affinity group g:
 This is the information about the set of nodes in the same group.
The data includes the roundtrip time estimate, the heartbeat
count, etc.
 Contacts of all other affinity groups :
 This is the information about a small constant number of nodes in
all other groups.
 Filetuples:
 This is the intra-group index about the set of files whose
homenodes are in the same
 affinity group.
 A file tuple consists of a file name and the IP address of the file’s
homenode. A heartbeat count is also associated with a file tuple.

37

 The total number of routing table entries per
node is-
N/k+c*(k- 1 )+F/k

Where,
N: refers to the total number of nodes,
c: the number of contacts per group,
F: total number of files in the system, and
k: the number of affinity groups

F is proportional to N and c is fixed

38

 To look up a file f, the querying node A in the group G
hashes the file to the file’s belonging group G'.
 If G' is the same as G, the query is resolved by checking the
node A’s local data store and local intra-group data index.
 Otherwise, A forwards the query to the topologically closest
contact in group G'.
 On receiving a query request, the contact in the group G'
searches its local data store and local intra-group data index.
The IP address of f’s homenode is then returned to the
querying node directly.
 In case of a file lookup failure, the querying node retries
using different contacts in the group G’
 Using a random walk in the group G' , or a random walk in
the group G.

39

 Coral in is an indexing scheme. It does not dictate how to
store or replicate data items.
 Objectives of Coral are to avoid hot spots and to find
nearby data without querying distant nodes.
 A distributed sloppy hash table (DSHT) has been proposed
to eliminate hot spots.
 In DHT, a key is associated with a single value which is a
data item or a pointer to a data item.
 In a DSHT, a key is associated with a number of values
which are pointers to replicas of data items.
 DSHT provides the interface: put(key, value) and get(key).
 put(key,value) stores a value under a key.
 get(key) returns a subset of values under a key.

40

 When a file replica is stored locally on a node A,
 The node A hashes the file name to a key k and inserts a
pointer nodeaddr (A’s address) to that file into the DSHT
by calling put(k,nodeaddr).
 To query for a list of values for a key k, get(k) is forwarded
in the identifier space.
 Coral organizes nodes into a hierarchy of clusters and puts
nearby nodes in the same cluster.
 Coral consists of three levels of clusters. In the lowest-level, Level
2, cover peers located in the same region and have the cluster
diameter (round-trip time) 30msecs.
 Level 1, cover peers located in the same continent and have the
cluster diameter 100msecs.
 Level 0, is a single cluster for the entire planet and the cluster
diameter is infinite.

41

 DHT balance load among different nodes, but hashing
destroys data locality.
 The non-DHT P2Ps try to solve the problems of DHT P2Ps
by avoiding hashing.
 Hashing does not keep data locality and is not amenable to
range queries.
 Some non-DHT P2Ps are SkipNet, SkipGraph, and TerraDir.
 SkipNet is designed for storing data close to users.
 SkipGraph is intended for supporting range queries.
 TerraDir is targeted for hierarchical name searches.
 Searching in such systems follows the specified neighboring
 relationships between nodes.

42

 The overlay SkipNet in supports Content Locality & Path
Locality. by using a hierarchical naming structure.
 Content locality refers to the fact that a data item is stored close
to its users.
 Path locality means that routing between the querying node and
the node responsible for the queried data are within their
organization.

Figure: The peer name order and sample routing tables

43

 This chapter discusses various searching techniques in peer-
to-peer networks (P2Ps).
 Clearly, significant progress has been made in the P2P
research field. But still long way to go..
 First, good benchmarks need to be developed to evaluate the actual
performance of various techniques.
 Secondly, schemes amenable to complex queries supporting relevance
ranking, aggregates, or SQL are needed to satisfy the practical
requirements of P2P users
 Thirdly, security issues have not been addressed by most current
searching techniques.
 Fourthly, P2P systems are dynamic in nature. Unfortunately existing
searching techniques can not handle concurrent node join-leave
gracefully.
 Fifthly, good strategies are needed to form overlays that consider the
underlying network proximity.
 Sixthly,almost all existing techniques are forwarding-based techniques.
Recently, a study on non More effort is required to develop good non-
forwarding techniques and to compare non-forwarding techniques to
various forwarding techniques

45

Thank You
For Your Attention

46

20120412 searching techniques in peer to peer networks

More Related Content

What's hot (12)

Viewers also liked (10)

Similar to 20120412 searching techniques in peer to peer networks (20)

20120412 searching techniques in peer to peer networks

Editor's Notes