HPUX 와 오라클 DB 의 HA 아키텍처 구성

황 명 석

한국 HP / 2004.3.18

2004 MAR

MC Seminar

Agenda

1. Parallel Hardware Architecture

2. Oracle RAC 성능

3. Oracle RAC 와 hp Cluster Type

4. Oracle RAC 의 TAF

Parallel Hardware Architecture

Parallel Hardware Architecture

Parallel Hardware Architecture

•Tightly Coupled System•Loosely Coupled System•Massively Parallel System

Tightly Coupled System

• 다수의 CPU 가 memory 공유• 각 CPU 는 common Bus 를 통해 shared memory

access• Performance 는 Memory Bus 의 Bandwidth 에

의해 결정• SMP System (Symmetric Multiple Process)

Loosely Coupled Systems

• Loosely Coupled systems (cluster) 의 shared disk 는 하드웨어 레벨에서 구현 .

• Shared disk system 는 물리적으로 cluster node 간에 동일한 disk array 를 공유 .

Shared Nothing Implementations - MPP

• Shared Nothing Implementations - MPP– 각 노드는 하나 이상의 CPU 를 갖는다 .– 메모리는 노드간에 공유하지 않는다 .– High-speed interconnect 사용을 기반으로 한다 .– File 의 I/O request 는 user program 에게 투명한 high-

speed interconnect 를 통하여 remote disk 를 accessing 한다 .

Oracle RAC 의 성능Oracle RAC 의 성능

e-Business 시대의 RAC 필요성

e-Business 요구사항 높은 가용성 (High Availability) 높은 확장성 (High Scalability) 높은 서비스품질 (High Quality Of Services) 시장의 확대 (Extend Market – Globalization) DBA 생산성 향상 On-Line Management Business intelligence E-business 를 위한 개발 환경

Cache Fusion Architecture

• 완벽한 Cache Fusion– Cache-to-Cache

데이터 전송– 공유 캐쉬는 느린 I/O

작업을 줄인다 .– 데이터는 고성능

interconnect 를 통해 다른 노드에 직접 전송

• 확장성 제공

Oracle9i RAC Scalability

HP-UX 기반 Platform 에서의테스트 결과

Oracle RAC 와 hp Cluster Type

Oracle RAC 와 hp Cluster Type

HP ServiceGuard ClusterProtection Levels

• single cluster

• within single

data center

• automatic failover

• single cluster

• same site, separate bldgs

• automatic failover

• single cluster

• same city, separate sites

• automatic failover

• multiple clusters

• different cities, different countries

• “Push-button” failover

LocalCluster

Campus

Cluster

Metro Cluster

Continental

Cluster

FlexibilityFunctionality

Distance

Local Cluster supported with 9i RAC

• Distances• up to 500m with direct FC

• up to 11km with FC switches (500m to switch, 10km between switches)

• up to 25m with direct SCSI

• Single Cluster

• Automatic Failover

Data Center 1 Storage

RAC

Node A Node B

within single data center

Campus Cluster supported with 9i RAC

• Distances• up to 10km with Oracle9i RAC with FC switch/hub

• up to 100km only with MC/SG (extended CampusCluster) with DWDM

• Single Cluster

• Automatic Failover

• Software Mirroring with HP MirrorDisk/UX

• Only 2 node cluster supported with RAC & SLVM, and 4 nodes supported with RAC & Veritas CVM.

• For all storage types

Data Center 1 Data Center 2

Storage Storage

RAC

Node A

MirrorUX

Node B

MirrorUX

same site, separate buildings

Metro Cluster not supported with 9i RAC

• Distances for CA/SRDF• ESCON, ca. 43km• FC direct, ca. 500m• FC switches, ca. 10km• FC switches plus ultra long haul GBIC ca. 80km• DWDM, ca. 100km

• Single Cluster

• Rapid, automatic site recovery without human intervention

• Storage Hardware Mirroring

• separate arbitrator for split brain situations

Data Center 1 Data Center 2

XP/EMC XP/EMC

MC/SG

Node A Node B

Node CArbitrator

CA / SRDF

same city, separate sites

Continental Cluster supported with 9i RAC

• Distances•Unlimited distance between data centers!

• Separate Cluster

• Locate data centers at economically and/or strategically best locations

• “Push-button” failover across 1000s of km

different cities, different countriesData Center 1 Data Center 2

XP/EMC

RAC

Node A Node B

XP/EMC

RAC

Node A Node B

WAN

Oracle DataGuard supported with 9i RAC

•Distances• no limitation for Oracle DataGuard

• RAC cluster in one data center

• Changes get propagated to second data center

with Oracle DataGuard (formerly Automated Standby Database)

disaster tolarantData Center 1 Data Center 2

Storage

RAC

Node A Node B Node C

Storage

OracleDataGuard

Standby Database History 정리

• Oracle Version 7.3 custom Standby Databases– Current Log 반영 불가 (Only transferred arch. files)

• Oracle8i Automated Standby– Read only databases, Managed recovery– Remote archiving

• Oracle8i Data Guard– Automation: Single command switchover and failover– OPS and OFS support

• Oracle9i Data Guard– Built in zero Data Loss capability (Current log 반영 )– GUI interface integrated with OEM

• Oracle9i R2 Data Guard– Oracle9i Data Guard Logical Standby database– SQL apply instead of log– Query & Reporting enabled

Oracle RAC 의 Transparent Application Failover

Oracle RAC 의 TAFOracle RAC 의 TAF

page 27

What do you want to avoid?

•Database connections are stateful, simple network failover is insufficient

•Even after a fast database recovery clients were forced to exit their application and reconnect to the database•Very intrusive

•Work lost

page 28

Client Side Load Balancing

• Clients connect to instance using random method (uses address list in tnsnames.ora)

Node 1

rac2instancelsnr2

9iRACDatabase

rac1instancelsnr1

Node 2

Netw

ork

Client 1

Client 2

Client 3

Client n

page 29

Listener Load Balancing

• Listeners balance load using CPU/user load

Node 1

rac2instancelsnr2

9iRACDatabase

rac1instancelsnr1

Node 2

Netw

ork

Client 1

Client 2

Client 3

Client n

page 30

Listener Load Balancing

• Listeners balance load using CPU/user load

Node 1

rac2instancelsnr2

9iRACDatabase

rac1instancelsnr1

Node 2

Netw

ork

Client 1

Client 2

Client 3

Client n

page 31

Connect Time Failover

• Automatically retries the connection (uses the next entry in the address list in tnsnames.ora)

Client

TNS

Node 1

rac2instancelsnr2

9iRACDatabase

rac1instancelsnr1

Node 2

page 32

TAF, Overview

• Oracle9i provides a high availability architecture that provides transparent client failover capability–Little or no user downtime –Applications and users are automatically and

transparently reconnected to another system

–Applications and queries continue uninterrupted–Login context maintained

ComputerComputer

AAComputerComputer

BBComputerComputer

AAComputerComputer

BB

Node A in an RAC cluster fails, users

are migrated

page 33

Characteristics of TAF

• TAF protects or fails-over:• Client/server connection

• User session state

• Active cursors (select statements) that have begun to return results

• Not failed over: - Active update transactions - PL/SQL server-side package variables

page 34

Three “Levels” of TAF Functionality

• TYPE=SESSION, METHOD=BASIC (Login Failover)•Client is automatically logged into surviving node of cluster

• TYPE=SELECT (Statement Failover)

•Node failure occurs during query

•Client fails over to a surviving node and is logged in

•Query replayed on surviving node but only rows not returned

during the original query execution are returned

• METHOD=PRECONNECT (“Fast” Session Failover)

•Client connected to two instances at session establishment

•Avoid impact of “login storm” during failover to surviving node in

Real Applications Cluster or Oracle Parallel Server

page 35

TAF Select Failover

• Failover allows the application to continue execution or fetching

• Leverages Oracle’s multi-versioning read consistency to ensure results are identical

SELECT * FROM emp; SELECT * FROM emp;

Instance 2

Client

empno name

7369 Smith7499 Allen7521 Ward7566 Jones7654 Martin7698 Blake

Instance 1

DB

empno name

7369 Smith7499 Allen7521 Ward7566 Jones7654 Martin7698 BlakeRows Partially Returned

When Failover Occurred

Continues ReturningRemaining Rows from here