Keyword Search on External Memory Data Graphs Bhavana Bharat Dalvi, Meghana Kshirsagar, S. Sudarshan PVLDB 2008 Reported by: Yiqi Lu

Background: Graph Model  Direct graph model for data

Background: Answer Tree Model  Answer Tree  Keyword Query

Background: score function  A function of the node score and edge score of answer tree  Several score models have been proposed.

Background: keyword search  Input: keywords, data graph  Output: top-k answer trees  Algorithm:  first  looking up an inverted keyword index to get the node-ids of nodes  Second  a graph search algorithm is run to find out trees connecting the keyword nodes found above. The algorithm finds rooted answer trees, which should be generated in ranked order.

Example: backward expanding search  For each keyword term k i  First find the set of nodes S i that contain keyword k i  Run Dijkstra SP algorithm which provides an interface to incrementally retrieve the next nearest node  Traverses the graph  To find a common vertex from which a forward path exists to at least one node in each set S i  Then the answer tree’s root is the common vertex and the keywords are leaves

Background: external memory search  Run search algorithm on an external memory graph representation which clusters nodes into disk pages  Naïve migration will lead to poor performance  keyword search algorithms designed for in-memory search access a lot of nodes, and such node accesses lead to a lot of expensive random IO when data is disk resident.

Background ： 2-level graph  Clustering parameters are chosen such that supernode graph fits into the available amount of memory

Background: 2-phase search algorithm

This algorithm lack consideration of time locality

multi-granular graph structure  This paper proposes a multi-granular graph structure to exploit information present in lower-level nodes that are cache-resident at the time a query is executed

MG graph  a hybrid graph  A supernode is present either in expanded form (all its innernodes along with their adjacency lists are present in the cache)  Or unexpanded form (its innernodes are not in the cache)

several types of edges

Supernode answer Pure answer

ITERATIVE EXPANSION SEARCH  Explore phase: Run an in-memory search algorithm on the current state of the multi-granular graph (the multi- granular graph is entirely in memory)  Expand phase: Expand the supernodes found in top-n results of the (a) and add them to input graph to produce an expanded multi-granular graph

ITERATIVE EXPANSION SEARCH

 the stopping criterion ：  The algorithm stops at the iteration where all top-k results are pure.  node-budget heuristic:  Stop search when

ITERATIVE EXPANSION SEARCH  A assumption: the part of graph relevant to the query fits in cache  May fail in some cases  Query has many keywords or algorithm explores a large number of nodes  Have to evict some supernodes from the cache based on a cache replacement policy  some parts of the multi-granular graph may shrink after an iteration  Such shrinkage can unfortunately cause a problem of cycles in evaluation

ITERATIVE EXPANSION SEARCH  do not shrink the logical multi-granular graph, but instead provide a “virtual memory view” of an ever-expanding multi-granular graph.  maintain a list, Top-n-SupernodeList, of all supernodes found in the top-n results of all previous iterations.  Any node present in Top-n-SupernodeList but not in cache is transparently read into cache whenever it is accessed.

INCREMENTAL EXPANSION SEARCH  Iterative Expansion algorithm restart search when supernodes are expanded  This can lead to significantly increased CPU time  Incremental expansion algorithm updates the state of the search algorithm

Take BES as example

Heuristics to improve performance  stop-expansion-on-full-cache  Intra-supernode-weight heuristic  We define the intra-supernode weight of a supernode as the average of all innernode → innernode edges within that supernode.

Experiment  Search Algorithms Compared  Iterative Expanding search  Incremental Expanding (Backward) Search with different heuristics  the in-memory Backward Expanding search  the Sparse algorithm from “Efficient IR-Style keyword search in relational databases”  A naive approach to external memory search would be to run in-memory algorithms in virtual memory.  we have implemented this approach on the supernode graph infrastructure, treating each supernode as a page

Data sets  DBLP 2003  IMDB 2003  Cluster using EBFS technique  Default supernode size is 100 innernodes corresponding to an average of 7KB on DBLP and 6.8KB on IMDB  Supernode contents were stored sequentially in a single file, with an index for random access within the file to retrieve a specified supernode.

Data sets

Clustering result

Cache Management  3GB RAM, and a 2.4GHz Intel Core 2 processor, and ran Fedora Core 6  All results are taken on a cold cache.  Force linux kernel to drop page cache, inode cache and dentry cache  By excuting sync(flush dirty pages back to disk) then excuting echo 3 > /proc/sys/vm/drop_caches

Queries

Experimental Results  first implemented Incremental search without any of the heuristics  did not perform well, and gave poor results, taking unreasonably long times for many queries.  results for this case not presented  two versions of Incremental expansion, one with and one without the intra-supernode-weight heuristic

the intra-supernode-weight heuristic reduces the number of cache misses drasti cally without significantly reducing answer quality.

Comparison With Alternatives