Oracle RAC and Linux in the real enterprise

1
Oracle RAC and Linux in the real enterprise

• October, 02
• Mark Clark
• Director
• Merrill Lynch Europe PLC
• Global Database Technologies

2
Agenda

• The challenge
• Current solutions
• DBMS Utopia at Merrill Lynch
• Selected architecture
• Selected partners
• Program history
• Results and conclusions
• The managed service

3
The Challenge and Mission

• Access to Oracles renowned availability,
  integrity, performance and scalability at a
  commodity server price point

4
Current solutions

• Risc based UNIX platforms
• Strengths
• Well understood and integrated into ML
  infrastructure
• Full and complete ISV support
• Robust and mature
• Low risk technically and financially
• Mass acceptance

5
Current solutions (cont.)

• RISC based UNIX platforms
• Weaknesses
• Platform costs moderately high
• Large compute platforms especially so
• Limited competition in server marketplace
• Emerging technologies slow to market
• Closed source

6
DBMS Utopia at Merrill Lynch

• Fully managed service
• Actual usage-based charging model
• Delivery at commodity prices
• Flexible capacity model
• High availability at Day 1

7
Selected technologies

• Oracle / RAC
• A market leader of database management tools and
  products with new high availability/scalability
  options (RAC).
• Intel
• Commodity server architecture, well understood at
  Merrill Lynch in the SQL/server domain.
  Compelling technology roadmap
• Linux
• Rapidly developing open source OS with a strong
  technology partnership with Oracle and Intel.
  Leverages the Intel architecture. Just another
  UNIX like OS.

8
Selected partners

• Oracle
• Early access to certain product functionality and
  new features. Fast track to engineering and
  performance group
• Hewlett Packard
• Current Intel server platform for Merrill Lynch
  Europe and provider of server hardware for all
  testing (formerly Compaq)
• Intel
• Provision of customer solutions centres and
  skills / resources for the offsite proof of
  concept

9
Program history

• Key milestones
• November 2001 (Phase 1)
• IA32, Oracle 9.0.1, Red Hat 7.2, Raw partitions
• March 2002 (Phase 1b)
• Two node production RAC deployment
• October 2002 (Phase 2)
• IA32, Oracle 9.2.0, Red Hat 2.1 AS, OCFS

10
Program Results - Phase 1

• November 2001 (Phase 1) POC Build
• IA32, Oracle 9.0.1, Red Hat 7.2, Raw partitions
• Functional
• Oracle worked
• Performance
• More performant and lower cost comparable to
  RISC/UNIX server of similar configuration

11
Program Results - Phase 1 (cont.)

• Gaps
• Non-certified OS/Oracle combination
• Infrastructure and ISV alignment
• Unmanageable RAW / No OMF support
• No cluster filesystem
• No SAN integration
• Backup / Recovery tools
• Performance management
• Monitoring / Alerting

12
Program Results - Phase 1b

• March 2002 (Phase 1b) - POC Build
• IA32, Oracle 9.0.1, Red Hat 7.2, Raw partitions
• Functional
• Oracle worked
• Performance
• More performant and lower cost comparable to
  RISC/UNIX server of similar configuration

13
Program Results - Phase 1b (cont.)

• Risks and mitigation

14
Program results - Phase 2

• October 2002 (Phase 2)
• The GA Build
• IA32, Oracle 9.2.0, Red Hat 2.1 AS, OCFS
• 4 Node RAC configuration
• 32 Processor
• 16GB memory
• EMC Clarion storage array
• Brocade switched SAN fabric
• Gigabit cluster interconnect with redundant
  switches
• Single instance per node (today)

15
Program results - Phase 2 (cont.)
16
Program results - Phase 2 (cont.)

• Functional
• It still worked
• Performance / Stress testing
• Server side Oracle CPU stress test
• I/O stress test (Direct path load)
• Client/Server realworld stress test
• Critical failure condition testing
• Unexpected power failures
• Interconnect failure (single path)
• SAN connectivity failure

17
Program results Phase 2 - (cont.)

• Performance / Stress testing
• Server side Oracle CPU stress test
• Used to prove CPU scalability
• Open source benchmarking tool based broadly on
  TPC-C Order-Status Stock-Level transactions -
  orabm sourced from http//www.dbcool.com
• In memory database (no I/O stress) - 300MB SGA
  total
• orabm allows rapid deployment and results
  gathering - ie low cost

18
Program results Phase 2 - (cont.)

• Results of best single session tps
• (Oracle CPU stress test)

19
Program results Phase 2 - (cont.)

• Performance / Stress testing
• I/O stress test (Direct path load)
• Used to prove I/O scalability
• Oracle direct path loader was used in a
  multi-stream parallel load activity
• Fixed format input data loading the TPC-C
  CUSTOMERS table
• Tests executed in a single node and multi node
  configuration

20
Program results Phase 2 - (cont.)

• Results of I/O stress test

Note 6 load streams per node (ie 12 streams 2
nodes)
21
Program results Phase 2 - (cont.)

• Performance / Stress testing
• Client/Server realworld stress test
• Used to prove transaction throughput compared to
  existing Merrill Lynch systems
• Using Quest Benchmark Factory
• Loading generated 10x change volumes compared to
  heaviest loaded ML EMEA database

22
Program results Phase 2 - (cont.)

• Critical failure condition testing
• Unexpected power failures
• Removed the power cable mid-processing
• No outage
• Interconnect failure (single path)
• NIC failover by Software
• Completely transparent to Oracle
• No outage
• SAN connectivity failure
• Removal of HBA cable mid processing
• No outage

23
Phase 2 - summary

• Gap resolution for DBMS Utopia

24
The managed service - the future

• Based on the 2002 Phase 2 configuration and
  products
• Flexible capacity model
• Rapid time to solution provision
• With an actual compute-based billing model
• Due for launch in December 2002