Federating Scientific Data

1
Federating Scientific Data

• Tanu Malik
• Dept. of Computer Science
• Johns Hopkins University

2
New trends in scientific research

• Explore the same scientific phenomena in related
  sub-disciplines
• In Astronomy, the same sky is observed in
  different electromagnetic spectra
• In Biology, cures for the same disease is
  searched in genetics, physiology and molecular
  biology
• Trend is important for new discoveries to take
  place
• Correlation between new discoveries and the
  number of connections between fundamental
  properties
• Number of connections increase by federating data
  from sub-disciplines

3
Issues in building federations

• Systems in scientific sub-disciplines are
  autonomous and thus heterogeneous
• Sub-disciplines make independent choices in
  hardware and software
• Federations can become non-inter-operable
• Scientific sub-disciplines are data-intensive
• Data acquisition governed by Moores law
• Federated tasks (queries) compare, merge and join
  large data-sets from
  multiple sub-disciplines
• Efficient execution of tasks
• By optimizing individual tasks before execution
• By minimizing volume of data transfer from all
  tasks
• Collection of task-specific statistics might be
  essential for efficient execution
• Need for effective solutions to allow for
  efficient federation,
  management, and correlation of data

4
Astronomy Our representative science

• Astronomy needs federations
• Currently, more than 100 independent surveys
  published on the Internet
• Typically each contains 10-100 million objects
• Different wavelengths, different properties, but
  same objects
• Several questions for finding similar objects
• Real, high-dimensional data
• Astronomy data has no commercial value
• No privacy concerns. Free for sharing
• Great for experimenting

5
Our contributions

• A federation of astronomy archives The SkyQuery
• Architecture, specification of federated queries,
  algorithms for optimizing federated queries
• A caching framework
• Consists of algorithms that minimize network
  traffic. Algorithms build upon economic
  principles
• Estimating query result size
• A model that exploits data mining techniques to
  estimate result size of federated queries

6
The SkyQuery An astronomy federation

• Ability to query data across archives.
• Typical Query Find all objects near the Pole
  star that were observed by archives measuring
  infra-red wavelengths but not by archives
  measuring radio wavelengths
• This is a distributed probabilistic spatial join
  query as there is error associated with the
  recorded location of an object.
• Federation is necessary to run this query as
  infra-red and radio archives are independently
  managed
• Queries are to be optimized as the near
  criteria can potentially transfer thousands of
  objects over the Internet

7
SkyQuery Architecture
8
SkyQuery status

• Uses Web services standard for federation
• Currently a federation of 30 sites
• Is expected to grow to 120 sites soon
• Actively used by astronomers all over the world
• The spin-off is OpenSkyQuery that provides
  intuitive interfaces and
  fully-developed language
  for cross-match specification
  
• http//www.openskyquery.net (User site)
• http//www.skyquery.net (Research and
  Development site)

9
Bottleneck in SkyQuery

• In January 2004, a participating site of SkyQuery
  resulted in
• Total number of queries 26,667
• Total number of object accesses 45,154
• Total bytes on the network 1200 GB ( More than 1
  TB!)
• Total number of unique object accesses 3890
• Total size of objects that were accessed 90 GB
• SkyQuery does not respect good network
  citizenship
• Locality in object accesses can be translated
  into effective caching methods

10
Naïve caching is dangerous

• Risks of caching
• Caching may transfer unfiltered data
• May transfer columns or tables across network
• Caching moves data to programs/queries
• Caching may reduce parallelism
• May move queries from many servers to few caches

• Benefits of a federation
• Filters data prior to transfer
• Selection and aggregation queries reduce data
  size
• Moves the program (fewer bytes) to the data (more
  bytes)
• Executes queries in parallel
• Experiment scales with the number of federated
  sites

11
Bypass-Yield An effective caching method

• Distinguish between queries
• To bypass
• To satisfy in cache
• Bypass queries
• Sent directly to database servers
• Caching them does not save network traffic
• Allows for benefits of federation
  moving-queries-not-data,
• parallelism, and data filtering
• Queries satisfied in cache
• Save network traffic
• Defines yield of a query to
• Distinguish queries
• Helps decide accessed data objects to load from
  servers to cache
• Caching these data objects saves network traffic
• Measures the network savings rate per unit of
  cache space used

12
Network flow in a Bypass-Yield Cache
DL load cost, DB bypass cost Minimize WAN cost
on (DL DB)
13
The BYC metric

• Let si be the size and fi be the fetch cost of
  object oi
• Let yi,j be the size of the query-result of the
  jth query that accesses oi
• Let pi,j be the probability of occurrence of the
  jth query above
• Byte-Yield Hit Rate (BYHR) for object oi defined
  as
• BYHR has two components
• Expected benefit of caching oi due to yield
• Scaled by the cost/size ratio

14
BYC algorithms make economic decisions

• Rate-Profile An optimistic algorithm
• Predicts object access patterns from workload
• Combines recency and frequency to measure
  probability of access for objects in and outside
  cache
• Evicts objects with lowest rate of savings.
  Optimistically, loads objects whose rate of
  savings are increasing
• OnlineBY A defensive algorithm
• Makes no assumptions about workload
• Loads an object into the cache when network
  traffic equal to the size of object has been
  incurred
• Maintains the cache according to an web caching
  algorithm (GDS)
• Is lg2(k) competitive, k is the ratio of size of
  cache, to the smallest object in cache
• SpaceEffBY A randomized algorithm
• Maintains no state about workload
• Loads objects with probability equal to
  yield/size of object

15
Current researchYield estimation

• Selectivity Estimation
• Estimating the number of tuples that satisfy a
  query
• Accurate estimates can help in
• For choosing best plans in query optimization
• As feedback to users
• Resource consumption, load balancing
• For BYC to achieve maximum network savings
• Accurate selectivity estimation is crucial to BYC
  performance
• Governs load and evict decisions

16
Selectivity estimation in BYC

• Common methods to estimate selectivity
• Random sampling, histogram, wavelet based
  compressions
• A caching environment introduces new requirements
  on the selectivity estimation problem
• Caches are close to clients not servers often
  have no access to databases
• Cache is a constrained resource in terms of
  storage
• Caching systems are adaptive and online

17
Template example

• Assume there are large no of queries to the
  following query template
• SELECT order date,required date FROM Order
• WHERE freight OP
  VALUE
• VALUE varies over the domain A,B. A,B are real
  values. Operator OP comes from the set lt,gt

18
Freight distribution

• Cumulative frequency distribution for freight
  attribute for OP lt
• Needs to known precisely to estimate selectivity
  of all queries

19
Template example

• Assume there are large no of queries to the
  following query template
• SELECT order date,required date FROM Order
• WHERE freight OP
  VALUE
• VALUE varies in the range M,N. M,N are real
  values. Operator OP comes
  from the set lt,gt

• The feature vector consists of OP,VALUE
• Consider feature vectors over all such queries.
  Assume
  selectivity is known for some queries (say 25),
  in addition, say the
  selectivity can be classified into 3 classes
  low, medium and high

20
Estimation example

• Decision tree learns u, and v from query
  attributes
• Use linear regression in each class to obtain
  actual yield values

21
Properties of our solution

• Uses a query based approach
• Current methods use a data-based approach
• Detect templates in the workload
• Most scientific experiments adhere to query
  templates
• Very fast and compact
• Use decision trees and regression over templates
• Crude estimates for the bypass phase, accurate
  estimates for loading data
• Able to process complex queries
• Extracts feature vectors from templates

22
Future directions

• To improve query execution performance in the
  cache
• In the absence of indices and materialized views
• Dynamic physical design which reorganizes itself
  according to workload
• Open questions
• Finding the best file organization
• How to best represent queries in terms of
  database schema, without actually changing the
  physical organization?
• How to update changes to schema when the workload
  changes?
• Plan to work on the above problem till Spring of
  2006.

23
Acknowledgements

• Inside
• Comp. Science Randal Burns, Members of HSSL
• Phy Astro Alex Szalay, The SDSS Team
• JHU The GBO Committee
• Outside
• Microsoft Research Jim Gray
• Univ. of Notre Dame Amitabh Chaudhary, Nitesh
  Chawla
• Carnegie Mellon Anastassia Ailamaki, Stratos
  Papadomanolakis

24
(No Transcript)
25
Web services

• SkyQuery uses Web services for interoperability
• Use of Internet standards
• Communication Protocol HTTP
• Message Exchange Model
• Simple Object Access Protocol(SOAP) with
  eXtensible Markup Language(XML) encoding
• Describing, Defining and Discovering Web
  Services
• Web Services Description Language (WSDL),
  Uniform Discovery, Description and Integration
  (UDDI) of Web Services

26
The object cross-match

• Matching objects across archives, if they
  correspond to the same astronomical body.
• Given A set of observations, one from each
  archive.
• Question Do these observations refer to the same
  astronomical body?
• Answer
• Easy, if positions can be measured precisely
• the X Match answer is deterministic.

27
Probabilistic Cross Match

• Measured positions have an error due to
  instrument inaccuracies.
• This error follows a Gaussian distribution
  (and is known for each archive)
• Given A set of observations, one from each
  archive
• Question What is the probability that these
  observations refer to the same astronomical body?
• Algorithm ?
• Probabilistic Cross Match

28
Computing the probability of a X Match

• To compute the probability of a X Match, we
  assume a position for the real body.
• The probability of a X Match varies with this
  considered position of the real body.
• We consider that position for which probability
  is highest.
• This position is a kind of weighted mean.
• Probability at this position can be represented
  by a circular region about mean.
• Our X Match algorithm reports those sets of
  observations
• probability gt threshold

29
The probabilistic cross-match
30
Background Page Caching

• Fixed size objects/pages, different fetch cost
• Cache hit is equivalent to an entire page being
  accessed
• Caches pages that have high fetch cost
• Used in operating systems

P6
P6
Goal Minimize total fetch cost of pages
31
Background Object Caching

• Variable size objects, different fetch cost
• Cache hit is equivalent to accessing an entire
  object
• Caches objects that have high cost/size ratio.
• Used in proxy web caching

O8
08
Goal Minimize total fetch cost of objects
32
Bypass-Yield Caching

• Objects are of variable size
• For each object, cost of access varies
• Cache hit implies fetching an entire object, part
  of an object, or an aggregate computed over an
  object

Client Requests
Q8
Goal Minimize total network cost (load and
bypass)
33
Rate-Profile

• Estimating probability of object access to
  estimate savings
• By characterizing workload
• For objects in cache
• Estimates probability as frequency-counts over
  cache lifetime
• Computes a rate of network savings
• Every in-cache access increase the rate in
  proportion to query yield
• Every time step (query access) decays rate

34
Rate-Profile

• For objects outside cache
• Separates queries accesses for an object into
  clusters (episodes)
• For each episode maintains frequency-counts over
  episode length
• Maximum rate in each episode is the maximum
  benefit of caching An
  optimistic decision
• Ages maximum benefits in each episode by weight
• Computes the expected rate of network savings

35
Rate-Profile Algorithm

• Discount object load cost from expected savings
• Penalize object for the network cost incurred in
  loading an object into cache
  
• Load Vs Bypass
• Compares discounted expected rate of network
  savings of an object not in the
  cache with the rate of network savings
  of all objects in the cache
• If greater, load. Else, bypass
• Eviction
• Evict object(s) with lowest current savings

36
OnlineBY

• Formulates the bypass-yield caching problem as a
  combination of
• Rent-or-buy problem, and
• Object caching problem
• The bypass Vs. load decision A defensive
  decision
• Bypasses queries (rents) until total bypass
  traffic exceeds fetch cost
• Then, loads (buys) an object
• Eviction decisions and in-cache policies
• Cache is maintained according to an algorithm for
  object
• caching problem
• This is O(lg2k) competitive
• k size of cache / size of smallest object

37
Randomized Ski Rental

• Randomized version of OnlineBY
• Does not store the amount of traffic bypassed
• Load decision A random decision
• Load an object with probability equal to the
  ratio of the yield of the
  current query to the fetch cost
• Requires no metadata for making load Vs bypass
  decision
• Reduces total amount of meta data required
• Performs well in practice ( no bounds yet )

38
Network Cost of a Trace (columns)

• Rate-Profile dynamically chooses between no cache
  and inline cache
• Rate-Profile compares well with an optimal static
  cache
• Other algorithms show similar results

39
Algorithm Cost Analysis

• Rate-Profile is better than OnlineBY in column
  caching
• SpaceEffBY always lags behind
• In all algorithms, column caching is better than
  table caching

40
Schema Locality

• Schema locality implies reuse of schema elements
  
• than data elements
• Both tables and columns show heavy and long
  lasting
• periods of reuse

41
Containment

• It is imperative for workload to show both
  containment and
• locality for query caching to be viable (Luo et
  al)
• Determining exact query containment is NP
  complete
• (Chandra and Merlin)
• Few objects experience reuse

42
Architecture

• Status
• We are currently implementing the above system
  within BYC.
• We currently have preliminary experiments prove
  the efficacy of the above model.

43
Results(1) Range Query
44
Result(2) User-defined Function Queries
45
Result(3) Index Queries