Exploiting locality for scalable information retrieval in peer-to-peer networks

1

• Exploiting locality for scalable information
  retrieval in peer-to-peer networks
• D. Zeinalipour-Yazti, Vana Kalogeraki, Dimitrios
  Gunopulos
• Manos Moschous
• November, 2004

2
Issues in p2p networks

• Content based / file identifiers information
  retrieval.
• Dynamic networks (ad-hoc).
• Scalability (global knowledge).
• Query messages (flooding network congestion).
• Recall rate.
• Efficiency (recall rate / query messages).
• Query Response Time (QRT).

3
IR in pure p2p networks

• BFS technique
• Each peer forwards the query to all its neighbors
• Simple
• Performance
• Network utilization
• Use of TTL
• RBFS technique
• Each peer forwards the query to a random subset
  of its neighbors
• Reduce query messages
• Probabilistic algorithm

4
IR in pure p2p networks

• gtRES technique
• Each peer forwards the query to some of its peers
  based on some aggregated statistics.
• Heuristic The Most Results in Past (for the last
  10 queries).
• Explore..
• The larger network segments.
• The most stable neighbors.
• ! (The nodes which contain content related to the
  query.)
• gtRES is a quantitative rather than qualitative
  approach.

5
The intelligent search mechanism (ISM)

• Main Idea Peers estimate for each query, which
  of its peers are more likely to reply to this
  query, and propagates the query message to those
  peers only.
• Exploit the locality of past queries.
• Some characteristics
• Entirely distributed (requires only local
  knowledge).
• Scales well with the size of the network.
• Scales well to large data sets.
• Works well in dynamic environments.
• High recall rates.
• Minimize the communication costs.

6
Architecture (ISM) (1/4)

• Profiling structure
• Single queries table
• LRU policy to keep the most recent queries
• Table size is limited ? good performance

7
Architecture (ISM) (2/4)

• Query Similarity function (cosine similarity)
• Assumption A peer that has a document relevant
  to a given query is also likely to have other
  documents that are relevant to other similar
  queries.
• Qsim Q2 ?0,1
• L the set of all words appeared in queries
  1,1,1,1
• q1,1,0,0
• qi1,0,1,0

8
Architecture (ISM) (3/4)

• Peer ranking (Relevance Rank)
• Pi each peer.
• Pl the decision-maker node.
• a allows us to add more weight to the most
  similar queries.
• S(Pi, qj) the number of results returned by Pi
  for query qj.

9
Architecture (ISM) (4/4)

• Search Mechanism
• Invoke RR function.
• Forward query to k (threshold) peers only.

10
Experiments

• Peerware A distributed middleware infrastructure
• GraphGen generates network topologies.
• dataPeer p2p client which answers to boolean
  queries from its local xml repository(XQL).
• SearchPeer p2p client that performs queries and
  harvest answers back from a Peerware network
  (connect to a dataPeer and perform queries).

11
Experiments - DMP

• If node Pk receives the same query q with some
  TTL2, where TTL2gtTTL1 we allow the TTL2 message
  to proceed.
• This may allow q to reach more peers than its
  predecessor
• Without this fix the BFS behaviour is not
  predictable and therefore is not able to find the
  nodes that we were supposed to find.
• Our experiments revealed that almost 30 of the
  forwarded queries were discarded because of DMP.
• The experimental results presented in this work
  are not suffering from DMP.
• This is the reason why the number of messages is
  slightly higher (30) than the expected number
  of messages.
• The total number of messages should be
  for n nodes each of which with a degree
  di.

12
Experiments-DMP

• Query examples
• A set of 4 keywords
• 1 keyword gt 4 characters

• Random Topology
• Each vertex selects its d neighbors randomly.
• Simple.
• Leads to connected topologies if the degree d gt
  log2n.

Query
1 AUSTRIA INTERVENE DOES DOLLAR
2 APPROVES MEDITERRANEAN FINANCIAL PACKAGES
3 AGREES PEACE NEW MOVES
13
Experiments (Set1)

• Reuters 21578 Peerware
• Random topology of 104 nodes (static) with
  average degree 8 (running on network 75
  workstations).
• Categorize the documents by their country
  attribute (104 country files - each for a node) -
  Each country file has at least 5 articles.
• Data Sets
• Reuters 10X10 set of 10 random queries which are
  repeated 10 consecutive times (high locality of
  similar queries) suits better to ISM.
• Reuters 400 set of 400 random queries which are
  uniformly sampled from the initial 104 country
  files (lower repetition).

14
Results (Set1) Reuters 10X10 (1/4)

• Reducing query messages
• ISM finds the most documents compared to RBFS and
  gtRES.
• ISM achieves almost 90 (recall rate) while using
  only 38 of BFSs messages.
• ISM and gtRES start out with low recall rate.
• Suffer from low recall rate.

15
Results (Set1) Reuters 10X10 (2/4)

• Digging deeper by increasing TTL
• Reach more nodes deeper.
• ISM achieves 100 recall rate while using only
  57 of BFSs messages with TTL4.

16
Results (Set1) Reuters 10X10 (3/4)

• Reducing query response time (QRT)
• 30-60 of BFSs QRT for TTL4 and 60-80 for
  TTL5.
• ISM requires more time than gtRES because its
  decision involves some computation over the past
  queries.

17
Results (Set1) Reuters 400 (4/4)

• Improving the recall rate over time
• ISM achieves 95 recall rate while using 38 of
  BFSs messages.
• During queries 150-200 major outbreaks occur in
  BFS.
• ISM requires a learning period of about 100
  queries before it starts competing the
  performance of gtRES.

18
Experiments (Set2)

• TREC-LATimes Preeware (random topology of 1000
  nodes static)
• It contains approximately 132,000 articles.
• These articles were horizontally partitioned in
  1000 documents (Each document contain 132
  articles).
• Each peer shares one or more of 1000 documents
  (replicated articles).

19
Experiments (Set2)

• Data Sets
• TREC 100 a set of 100 queries out of the initial
  150 topics.
• TREC 10X10 a list of 10 randomly sampled
  queries, out of the initial 150 topics, which are
  repeated 10 consecutive times.
• TREC 50X2 for which we first generated a set
  a50 randomly sampled queries out of the initial
  150 topics merged with a generated list of
  another 50 queries which are randomly sampled out
  of a.

20
Results (Set2) TREC100 (1/3)

• Searching in a large-scale network topology
• For TTL5 we reach 859 of 1000 nodes (BFS).
• For TTL6 we reach 998 of 1000 nodes at a cost of
  8500 m/q.
• For TTL7 we reach all nodes at a cost of 10,500
  m/q.
• ISM will not exhibit any learning behavior if the
  frequency of terms is very low.

21
Results (Set2) TREC 10X10 (2/3)

• The effect of high term frequency
• The recall rate will improve dramatically if the
  frequency of terms is high.
• ISM achieves higher recall rate than BFS (BFSs
  TTL5).
• After the learning phase of 20-30 queries it
  scores 120 of BFSs recall rate by using 4 times
  less messages.

22
Results (Set2) TREC 50X2 (3/3)

• The effect of high term frequency
• More realistic set, a few terms occur many times
  in queries and most terms occur less frequently.
• ISM monotonically improves its recall rate and at
  the 90th query it again exceeds BFS performance.
• gtRESs recall rate fluctuate and behave as bad as
  RBFS if the queries dont follow any constant
  pattern.

23
Experiments (Set3)

• Searching in dynamic network topologies
• Why network failures?
• Misusage at the application layer (shutdown PC
  without disconnecting).
• Overwhelming amount of generated network traffic.
• Because of some poorly written p2p clients.
• Simulate dynamic environment
• Total number of suspended nodes is no more than
  drop_rate.
• drop_rate is evaluated every k seconds against a
  random number r.
• If r lt drop_rate node will break all incoming and
  outgoing connections (for l seconds).
• In our experiments
• K60,000 ms and l60,000 ms.
• TREC-LATimes Peerware with the TREC 10X10 query
  set.
• drop_rate belongs to (0.0, 0.05, 0.1, 0.2)
• r is a random number which is uniformly generated
  in 0.0 .. 1.0)

24
Results (Set3) (1/3)

• BFS mechanism
• The increase of drop_rate decreases the number of
  messages.
• BFS does not exhibit any learning behavior at any
  level of drop_rate.
• BFS is tolerable to small drop_rates (5) because
  is highly redundant.

25
Results (Set3) (2/3)

• gtRES mechanism
• The increase of drop_rate decreases the number of
  messages.
• gtRES does not exhibit any learning behavior at
  any level of drop_rate.

26
Results (Set3) (3/3)

• ISM mechanism
• The increase of drop_rate decreases the number of
  messages.
• Quite well at low levels of drop_rate.
• Not expected to be tolerant to large drop_rates
  (The information gathered by the profiling
  structure becomes obsolete before it gets the
  chance to be utilized).

27
Extend ISM to different environments

• ISM mechanism could easily become the query
  routing protocol for some hybrid p2p environments
  (KaZaa, Gnutella).
• Super Peers form a backbone of infrastructure
  (long-time network connectivity).
• Regular Peers are unstable and less powerful.
• How could it work?
• Regular peer obtain a list of active Super peers.
• Connects to one or more Super peer and post
  queries.
• Super peer utilize the ISM mechanism and forward
  the query to a selective subset of its super peer
  neighbors.

28

• Thank you