CS632 Parallel Database Systems

1
CS632Parallel Database Systems

• 6 April 1999
• Rimon Barr

2
Background

• Database machine
• not economical
• parallelism
• Software database machine
• off-the-shelf components
• economy of scale

3
Software database architectures

• Shared memory
• processors on shared bus
• very fast, efficient communication
• not scaleable due to resource contention
• easy to program

4
Software database architectures

• Shared memory
• Shared disk
• private memory
• massively parallel architecture
• Thinking machines, Intel, N-cube, VAXcluster
• failed losses at both ends of the spectrum

5
Software database architectures

• Shared memory
• Shared disk
• Shared nothing
• highly scaleable
• processors with private memory and disks
• communication via interconnect

6
Parallel computation metrics

• Scaleup
• add processors in proportion to problem size
• Speedup
• add processors to fixed size problem

7
The Gamma Database Machine Project

• DeWitt, Ghandeharizadeh, Schneider, Bricker,
  Hsiao, Rasmussen.

8
History - DIRECT

• early database machine project
• showed parallelism useful for db apps
• flaws curtailed scalability
• shared memory
• central control of execution

9
Gamma hardware (v1.0)

• 17 VAX 11/750 processors
• 2 MB RAM per node
• 80 Mb/s token ring
• 8 x 333 MB Fujitsu drives
• Unix

10
Gamma hardware (v1.0) issues

• 2K DB pages due to token ring
• Unibus congestion (network faster?)
• corrected with a backplane card
• VAX obsolete
• 2MB with no virtual memory was tight

11
Gamma hardware (v2.0)

• iPSC/2 Intel hypercube
• 32 x386 processors
• 8MB of memory
• 330MB Maxtor drive / node (45KB cache)
• 8 routing modules
• 2.8 Mb/s
• full duplex, serial, reliable

12
Gamma software (v2.0)

• OS NOSE
• entire DB in one NX/2 process
• details renaming nodes
• 10 CPU used for copying
• excessive interrupts during I/O

13
Data storage

• Horizontal partitioning
• round-robin
• hashed
• range partitioned
• should partition relations based on heat

14
Components

• Catalog manager
• repository for db schema and meta-information
• Query manager
• one associated with each user
• Scheduler processes
• coordinates multi-site queries
• Operator processes
• executes single relational operator

15
Components
16
Query processing

• Ad-hoc and embedded query interfaces
• standard parsing, optimization, code gen.
• left deep trees only
• hash joins only
• at most two join operators active simul.
• split tables
• new relational operators

17
Split table

• Directs operator output to appropriate node

18
Query processing an example
19
Query processing an example
20
Algorithms selection

• start selection operator on each node
• semantic exclusion of nodes possible for hash and
  range partitioning
• throughput considerations
• transfer tuples in blocks
• one page read-ahead

21
Algorithms join

• Partition into buckets, join buckets
• Implemented
• sort-merge, Grace, Simple, Hybrid
• equi-joins
• centralized hash join (discuss)
• parallel hash join (discuss)

22
Algorithms aggregation

• compute partial results for each partition
• collect each group-by partition at single node,
  using hash function

23
Algorithms update

• standard techniques
• exception update to partitioning attribute may
  cause tuple to move

24
Concurrency control

• 2PL
• Granularity file and page
• Modes S, X, IS, IX, SIX
• local wait-for graphs
• centralized multi-site lock detector

25
Logging and recovery

• Log sequence number, LSN
• LSN
• Processor i directs log records at log manager (i
  mod M), where M log mgrs.
• standard WAL protocol
• ARIES

26
Node failure

• availability in spite of processor or disk fail
• mirrored disk (Tandem)
• interleaved declustering (Teradata)
• chained declustering
• (discuss)
• load redirection results in 1/n increase

27
Node failure - declustering
28
Node failure - load redirection
29
Performance experiments

• Selection
• Relation size
• Speedup
• Scaleup
• Join
• (similar)
• Aggregate
• Update (no recovery)

30
DiscussionThe Future of High Performance
Database Systems

• David DeWitt, Jim Gray.

31
Current day benchmarks

• TPC-C Online Transaction Processing
• TPC-D Decision Support
• IBM DB2 v5.2 TPC-D, 1000GB, NT

32
CS632Parallel Database Systems