Highly Available Database Systems

1
Highly Available Database Systems

• Seminar im WS 2005/2006
• Dependable Adaptive Information Systems
  (DAIS)Technische Universität Kaiserslautern
• Ou Yi

2
Agenda

• Introduction of Basic Concepts
• Three Types of Coupling
• DDBS PDBS
• Clustering
• HA in Practice

3
Availability Concept

• Why?
• Availability
• the fraction of the offered load that is
  processed with acceptable response times
• A MTTF / ( MTTF MTTR )
• The Number of Nines

4
HA and Multi-Computer System

• High Availability
• A ? 1
• downtime ? 0
• causes of downtime
• trouble-free components? redundancy!
• Multi-Computer System
• higher availability than single-computer system
• higher performance
• communication (network or shared HW)
• tight coupling, close coupling, loose coupling

5
Tight Coupling

• Characteristics
• shared main memory
• single copy of software
• Multiprocessor and SMP
• Pro Contra
• computing power
• communication
• - availability
• - extensibility (max 16)

6
Tight Coupling (2)

• Shared Everything (Shared Memory)
• a DBMS running on a multiprocessor
• multiprocessing with support of OS
• used as node architecture

tight coupling
7
Loose Coupling

• Characteristics
• interconnected through network
• independency
• Pro Contra
• error isolation
• extensibility (unlimited?)
• - communication

8
Loose Coupling (2)

• Shared Nothing
• independent computers (or nodes)
• database physically partitioned, logically
  unified
• each node runs a copy of DBMS, which has direct
  access only to its own partition

9
Loose Coupling (3)

• Shared Disk
• external storage is shared, all to all
• each node runs a DBMS, independent
• a TA can be completed locally
• no need for
• distributed query plan
• distributed commit
• required
• concurrency control
• buffer coherency control

10
Close Coupling

• Characteristics
• coupling component
• private main memory and software
• Pro Contra
• communication
• error isolation
• extensibility (max 32)
• - proprietary design
• Shared Data
• multi-computer DBS
• hybrid of close coupling and shared disk

11
Close Coupling (2)

• Shared Data (cont.)
• Parallel Sysplex Cluster
• shared disk close coupling
• Coupling Facility
• specialized SMP with large global main memory
• useful for global tasks
• Sysplex Timer
• global unique time
• e.g. global log file

12
Shared Disk vs. Shared Nothing

• Node Failure
• SD all the data still accessible to surviving
  nodes
• SN data of the failed node can not be easily
  accessed
• Increased Workload
• SD new nodes can be easily added and participate
  in query processing
• SN new nodes have no direct access to data,
  reallocation is expensive!

13
DDBS PDBS

• DDBS
• distributed database
• a collection of multiple, logically unified
  databases distributed over a computer network
• distributed DBMS
• manages the distributed DB
• maintains distribution transparency
• multiple DBMSs cooperating across sites (SN!)
• PDBS
• locally distributed database systems of the
  types shared nothing, shared disk, or shared
  everything

14
DDBS PDBS (2)

• Replicated Data
• across nodes, improve data availability
• improve performance, consistency?
• ideally one-copy equivalence
• ROWA (Read-Once/Write-All) algorithm
• one-copy equivalence
• - response time
• strict consistency is expensive!
• snapshot consistency
• snapshots, materialized views
• consistent in some point of time in the past
• acceptable for applications analysis, reporting,
  etc.
• periodical refreshing or triggered refreshing

15
DDBS PDBS (3)

• Disaster Recovery
• destroys complete computer center (site failure)
• keeping backup medium off-site
• time consuming
• server to be set up
• whole database to be retrieved
• loss of TAs
• online replication
• identical configuration
• highly up-to-date backup database
• quick takeover possible
• variations

16
DDBS PDBS (4)

• Online Replication (cont.)
• 1-safe
• asynchronous log transfer
• response time
• - loss of TAs
• very-safe
• distributed two-phase commit
• no loss of TAs
• - availability
• 2-safe
• backup involved in commit process in normal case
• primary independent from backup, unilaterally
  commit

17
Clustering

• Definition
• interconnecting of a group of computers so that
  they work together closely and it can be viewed
  as if it were a single computer
• special components for
• load distribution and failure detection
• architecture
• shared nothing, shared disk, shared data
• Cluster At Different Levels
• OS MS Windows NT/2000/2003, IBM AIX
• middleware C-JDBC
• application Oracle RAC (Real Application
  Clusters)

18
Clustering

• Purpose
• high-availability cluster
• fail-over switch-over
• backup is idle
• load balancing cluster
• all nodes are active
• Private Network
• used to detect node failure
• heartbeat status info. sent by the nodes to each
  otherat regular intervals
• network partition problem
• recommended to be redundant, fast and reliable

19
HA in Practice

• Mainframe (big iron)
• used by large companies
• mission critical applications and bulk data
  processing
• financial TA processing
• airline booking
• railway systems
• RAS, years without interruption
• IBM zSeries family
• z9-109 the most powerful
• with up to 54 configurable PUs (processor)
• and many HA features

z9-109 external view
20
HA in Practice - zSeries

• Memory-coherent SMP
• typical problem of SMP NUMA
• single large L2, uniform memory access, cache
  coherency

21
HA in Practice - zSeries

• PU Sparing Instruction Retry
• two spare PUs per server, protection against PU
  failure
• roles can be reassigned dynamically and
  transparently
• each instruction is executed two times
• detect soft error
• Modular Multi-book Design
• multiple books per server
• a book hosts PUs (12 / 16), memory and I/O
  connectors
• Concurrent Book Add (CBA)
• Enhanced Book Availability (EBA)

22
HA in Practice - zSeries

• Virtual Machine
• protected and isolated copy of the physical
  machine
• private address space, independency
• user illusion having a dedicated physical
  machine
• Hypervisor
• software which emulates the underlying physical
  machines architecture very efficiently (machine
  code)
• advantages
• error isolation
• overcoming hardware boundaries
• PR/SM of zSeries
• a pool of resources
• support up to 60 LPARs

23
HA in Practice - zSeries

• LPAR (Logical Partition)
• CPU resources
• zone (part of the physical main memory)
• I/O resources (statically assigned)

24
HA in Practice - DBMS

• System Failure
• database crash recovery clustering
• Data Failure
• storage failure
• RAID and (online) backupsDB2 backup db sample
  online to /dev3/backup
• human error
• Flashback technologyFLASHBACK TABLE account TO
  BEFORE DROPFLASHBACK DATABASE TO TIMESTAMP (...)
• Site Failure
• cross-sites online replication Oracle Data Guard
• ALTER DATABASE SET STANDBY DATABASE TO MAXIMIZE
• PROTECTION AVAILABILITY PERFORMANCE

25
HA in Practice - MAA

• Designing HA Architecture
• Oracle MAA (Maximum Availability Architecture)

26
Summary

• HA DBSs are multi-computer DBSs
• Basic architectures are the three types of
  coupling
• Error isolation and redundancy are effective and
  widely adopted approaches
• A promising market

27
Thank You!

• Technische Universität Kaiserslautern
• Ou Yi
• o_yi_at_informatik.uni-kl.de