Parallel Database System

1
Parallel Database System

• David Dewitt Jim Cray

presented by Ming Hao
2
Why parallel database

• dominance of Relational data model
• 1. Large uniform data record
• 2. Query can be decomposed into a bunch of
• relational operators. Each operator
  takes a
• relation as input and output a new
  relation
• 3. Indicate the built-in parallelism

3
1. pipelined parallelism streaming output of
one operator into the input of another
operator 2. partitioned parallelism
partitioned data and execution 3.
Inter-query parallelism OLTP
4
Hardware support available

• High speed network
• message passing based client-server operating
  system
• cheap and powerful PC/Workstation

5
Hardware architecture
1. Shared memory a. can not scale up to
lots of disks and processors network
bandwidth b. interference between
processors private cache does not solve
the problem
6
Hardware architecture
2. Shared disks a. same scale problem as
sharedM b. interference when updating data
7
Hardware architecture
3. Shared nothing a. linear scale up and
speedup b. less interference
c. exploiting commodity processors and
memory
8
Parallelism metrics

• Speedup
• small_system_elapsed_time
• big_system_elapsed_
  time
• scale up
• small_system_elapsed_time_on_small_proble
  m
• big_system_elapsed_time_on_big_problem

Speedup
9
Barriers to linear scaleup speedup

• Startup
• time to start parallel program
• interference
• critical section, synchronization,
  coherence
• skew
• load balance

10
Pipeline or Partitioning

• not very long operator chain
• pipeline not available for some operators
• aggregate
• skew

11
Data Partitioning

• Round-robin
• accessing data by sequential scan
• - frequently want to associatively access
  record

12
Data Partitioning

• hash partition
• accessing data by sequential scan
• frequently want to associatively access
  record
• - for clustering

13
Data Partitioning

• arrange partition
• accessing data by sequential scan
• frequently want to associatively access
  record
• clustering
• - data skew execution skew

14
Using existing sequential operators

• merge operator
• focusing data on one spot
• split operator
• used in multiple parallel stages
• flow control and buffering

15
Better algorithms

• Minimize data flow/ tolerate data and execution
  skew
• Join
• 1. Sort-Merge join nlog(n)
• 2. Hash -Join linear cost

16
summary

• Commodity components, not special hardware
• shared nothing architecture
• data partition, data flow
• only choice for some applications
• some remaining problems