Mariposa system

1
Mariposa system

• Witold Litwin

2
Basic goals

• WAN oriented DDBS
• Multiple sites
• e.g., 1000
• Scalable
• Locally autonomous
• Easy to evolve

3
Solution

• Traditional DBMS techniques
• a non-solution
• Economical paradigm rooted in
• Computer Life
• Litwin Wiederhold, 1990
• Contract nets
• Smith ?
• Implemented at Xerox Parc

4
DBMS architecture

• Object-relational data model
• Tables are fragmented
• hash, range, or user defined fragments
• replicas for better throughput
• fragments can split or coalesce (group)
• fragments can move among sites
• SDDSs could be used for the fragmentation

5
Manipulations

• SQL-3
• Queries are translated to operations on
  fragmentes
• An acceptable execution plan is produced
• perhaps with strides
• The plan is optimized through an economical
  tuning
• choose cheapest servers of subqueries

6
Architecture
Bidder
Name Server
Client query proc.
Bidder
Broker
Name Server
Bidder
Bidder
Bidder
Server query proc.
Server query proc.
Server query proc.
7
Query Processing Overview

• Rule based language for bidders, brokers and
  storage managers on the servers
• RUSH system (A.Sah al)
• brokers start query auctions
• sendout price / time curves
• in provided to sites by the network bank
• choose best bidders
• adjust query plans
• decreases the price as long as the total time
  remains acceptable

8
Bidders

• Propose prices times
• for CPU, IO, Memory for fragments
• Advertise their capabilities
• with name servers
• Select a, b from T -gt projections of T
• Select a, b from T where c '123' -gt selections
  and projections OK
• Advertise multioperational pricing plans
• volume reduction, coupons...

9
Pricing policy

• CPU IO c cost per time unit
• e (c, q) estimate of the subquery q cost
• l system current load l 1,2...
• final price p
• p e l
• Idle servers become work hungry
• Successful servers become expensive
• Work becomes naturally distributed

10
Memory services

• Similar rules as for queries, but
• Load l is the storage load factor
• The price is continuous per time unit
• Fragments can get sold among sites
• moves or copies
• The bidder bids only for queries
• to fragments it has permanently
• to temporary results it has from previous subquery

11
Bidding

• Expensive protocol
• auction out
• collect replies
• compute cheapest plan
• sendout queries
• Cheap protocol
• use only advertised capabilities
• issue purchase orders to selected bidders
• pay the bill (or go to court)

12
Name service

• Pretty classical naming schema thought
• name location discovery is dynamical
• is payable
• fragement moves are not synchronously posted to
  name servers
• as there are many name servers
• as in SDDSs for clients

13
Semantic Heterogeneity

• Name heterogeneity among sites resolved
• locally
• by name servers
• Data type and representation heterogeneity
• resolved using canonical representation

Integer
Integer
CR
CR
14
Conclusion

• Mariposa is an attempt to build large MDBSs
• based on human society management tools
• claimed more efficient at large scale than the
  traditional DBMS tools
• local autonomy
• no free lunch
• It is on-going, and 1st version implemented
• First results are successful
• One searches to further improve the economical
  tools

15
End