SmartSeer: Using a DHT to Process Continuous Queries Over PeertoPeer Networks

1
SmartSeer Using a DHT toProcess Continuous
Queries Over Peer-to-Peer Networks

• Authors
• Jayanthkumar Kannan
• Beverly Yang
• Scott Shenker
• Puneet Sharma
• Sujata Banerjee
• Sujoy Basu
• Sung-Ju Lee
• Presentation
• Colin Jiang

2
Motivation

• The academic world moves away from physical
  journals to online document repositories
• The ability to efficiently locate work of
  interest among the torrent of newly-generated
  papers will become increasingly important

3
SmartSeer

• Distributed preprint repository
• Supports rich DHT-based continuous (and
  instantaneous) queries
• Users register personalized continuous queries
  over the CiteSeer database and are alerted
  whenever papers that match their queries are put
  online

4
SmartSeers nonstandard design requirements

• SmartSeer supports a family of continuous queries
  that is far richer than just simple keyword
  searches
• SmartSeer is able of running on a loosely
  maintained group of unreliable machines spread
  across multiple organizations

5
Continuous vs Instantaneous Queries

• Latency is a prime design requirement for
  instantaneous queries but not for continuous
• In both types, queries are boolean conjunctions
  or disjunctions of search terms
• An inverted index is used in both cases for
  processing the queries
• Inverted Index
• ? Inverted List (per term in the document/
  query)
• ? IDs of all documents containing a given
  term

6
SmartSeer Architecture

• Push-based approach selected for minimizing
  cost
• The partition-by-keyword method was chosen
  among
• mirroring, partition-by-ID and
  partition-by-keyword
• Mirroring all documents and continuous
  queries are stored on all nodes ? not
  scalable to SmartSeers requirements
• Partition-by-ID continuous queries
  (documents) are partitioned
  among the nodes and a new document (query) is
  sent to all nodes ? not scalable
• Partition-by-keyword distributed index of
  the keywords in the continuous
  queries or documents using DHT and
  this index is partitioned among nodes
  based on the keyword

7
Insertion Functions

• Query Insertion
• A conjunctive query is stored at the node
  responsible for its most selective term ts (max
  load-balancing min wasted processing)
• The key of the query is defined to be the hash of
  ts
• Document Insertion (Send Query Method)
• Parsed into tokens (by the document insertion
  node- DN)
• Tokens are transformed to terms (by DN)
• All nodes responsible for at least one term in
  the document (query nodes- QN) are contacted via
  a document notification message
• QNs return the inverted list of all queries
  registered on that keyword
• DN does the matching and notifies the users

8
Send Query Bandwidth Consumption

• Instant Queries required bandwidth
• ß N I Wq Wd
• Continuous Queries required bandwidth
• Size Q(r) of the inverted list of queries
• Q(r) C FZ(r,a, Wq)
• Required bandwidth for the retrieval of Wd
  inverted lists Wd (Sr Z(r,a) Q(r))
• Bandwidth required to handle C continuous queries
  with documents arrival rate R per second
• ß RC F Wd

9
Send Query Alternatives

• Send Document
• When a new document is inserted, the entire
  document is sent to all the QNs storing the
  inverted lists for keywords in the document
• Each QN does the matching and notifies the users
• Term Dialogue
• The QN asks the DN about the presence of a set of
  keywords
• The DN responds with a bit vector specifying the
  presence of each term
• Shipping keywords can be much cheaper than
  shipping entire queries
• Bloom Filter
• The DN sends a bloom filter over all terms in the
  document to each of the QNs
• QNs can discard queries that have a term
  corresponding to a 0 in the bloom filter
• QNs may prune the set of queries stored in the
  inverted list to a smaller set

10
Alternatives Bandwidth Consumption

• Send Document
• SD d , d ? number of unique terms in the
  document
• Term Dialogue
• TD q , q ? number o unique terms across the
  queries
• Bloom Filter
• BF f(d) d Q
• Q Q0 S

11
Batch Notification

• Useful for median-scale systems
• Overlap in the nodes to which terms hash
• Use a clustered approach
• The DN will first find the QN responsible for
  each term
• The DN will send one document notification
  message to each unique node, along with the terms
  in the document that each node is responsible for
  

12
Evaluation (1/ 5)

• Written in Java
• Use of the libraries exported by Bamboo and
  OpenDHT
• SmartSeer runs over two document sets CiteSeer
  and Trec
• Workload of synthetically generated queries

13
Evaluation (2/ 5)

• Basic Comparison

14
Evaluation (3/ 5)

• Query Selectivity

15
Evaluation (4/ 5)

• Distribution of Terms

16
Evaluation (5/ 5)

• Batch Notification

17
Complex Queries

• Boolean Queries
• Negated Predicates SmartSeer registers queries
  with negated predicates on its non-negated
  predicates
• OR Queries These queries are registered on every
  term ? duplicated processing possibility ?
  rewrite the OR query to multiple queries such
  that only one of them is matched
• Boolean Queries decomposed to their disjunctive
  normal form ? sub-queries are joint via ORs and
  each sub-query is an AND query
• Nested Queries
• Nested Instantaneous Queries
• Execute first the sub-query
• ? retrieve a list of documents that satisfy this
  sub-query
• ? extract the appropriate attributes
• ? create on OR query which represents
  the translated version of the
  original sub-query
• ? execute this version
• ? retrieve a list of documents
  and use it in processing the original
  query
• Nested Continuous Queries
• The sub-query is translated as above
• ? both the original and the translated
  versions are registered (the
  translated for efficiency and the original for
  correctness)
• Negated terms are not allowed

18
Limitations on Expressiveness

• Similarity queries are not supported
• Instead, a query is registered on some keywords
  and when a new document is inserted, containing
  these keywords, we can check whether the new
  document is above a certain threshold of
  similarity with the query.
• Restrictions on the structure of the queries
• Queries with difficult (negation, range)
  predicates are not supported

19
SmartSeers Deployment

• 20 machines on Planetlab
• Average TCP throughput 10-15 Mbps
• A query corpus of about hundred thousand queries
  and a document corpus of 1000 documents
• Most inverted lists have very few queries
• The Send Document, Term Dialogue and Bloom Filter
  approaches can support 66,886, 78, 728 and 81,
  430 document insertions per day
• The throughput of the system was about 80,000 new
  documents per day

20
discussion
21
Bamboo DHT

• Bamboo DHT - A distributed hash table, or DHT, is
  a building block for peer-to-peer applications.
  At the most basic level, it allows a group of
  distributed hosts to collectively manage a
  mapping from keys to data values, without any
  fixed hierarchy, and with very little human
  assistance. This building block can then be used
  to ease the implementation of a diverse variety
  of peer-to-peer applications such as file sharing
  services, DNS replacements, web caches, etc.

22
Bamboo DHT(cont.)

• Bamboo is a either based on Pastry, a
  re-engineering of the Pastry protocols, or an
  entirely new DHT, depending on how you want to
  look at it. The term geometry is used to refer to
  the pattern of neighbor links in a DHT,
  independent of the routing algorithms or neighbor
  management algorithms used. Bamboo uses the
  Pastry geometry. It does not, however, use the
  same joining or neighbor management algorithms.
  Compared to Pastry, the algorithms are more
  incremental, a difference that allows Bamboo to
  better withstand large membership changes in the
  DHT as well as continuous churn in membership,
  especially in bandwidth-limited environments