By Brian on Nov 08, 2012
Oracle's SPARC T4-4 server delivered world record performance with subsecond response time on the Oracle OLAP Perf Version 2 benchmark using Oracle Database 11g Release 2 running on Oracle Solaris 11.
The SPARC T4-4 server achieved throughput of 430,000 cube-queries/hour with an average response time of 0.85 seconds and the median response time of 0.43 seconds. This was achieved by using only 60% of the available CPU resources leaving plenty of headroom for future growth.
Oracle OLAP Perf Version 2 Benchmark
4 Billion Fact Table Rows
|Users*||Response Time (sec)|
* Users - the supported number of users with a given think time of 60 seconds
Configuration Summary and Results
1 TB memory
2 x Sun Storage F5100 Flash Array (each with 80 FMODs)
Oracle Database 11g Release 2 (18.104.22.168) with Oracle OLAP option
The Oracle OLAP Perf Version 2 benchmark is a workload designed to demonstrate and stress the Oracle OLAP product's core features of fast query, fast update, and rich calculations on a multi-dimensional model to support enhanced Data Warehousing.
The bulk of the benchmark entails running a number of concurrent users, each issuing typical multidimensional queries against an Oracle OLAP cube. The cube has four dimensions: time, product, customer, and channel. Each query user issues approximately 150 different queries. One query chain may ask for total sales in a particular region (e.g South America) for a particular time period (e.g. Q4 of 2010) followed by additional queries which drill down into sales for individual countries (e.g. Chile, Peru, etc.) with further queries drilling down into individual stores, etc. Another query chain may ask for yearly comparisons of total sales for some product category (e.g. major household appliances) and then issue further queries drilling down into particular products (e.g. refrigerators, stoves. etc.), particular regions, particular customers, etc.
Results from version 2 of the benchmark are not comparable with version 1. The primary difference is the type of queries along with the query mix.
Key Points and Best Practices
Since typical BI users are often likely to issue similar queries, with different constants in the where clauses, setting the init.ora prameter "cursor_sharing" to "force" will provide for additional query throughput and a larger number of potential users. Except for this setting, together with making full use of available memory, out of the box performance for the OLAP Perf workload should provide results similar to what is reported here.
For a given number of query users with zero think time, the main measured metrics are the average query response time, the median query response time, and the query throughput. A derived metric is the maximum number of users the system can support achieving the measured response time assuming some non-zero think time. The calculation of the maximum number of users follows from the well-known response-time law
N = (rt + tt) * tp
where rt is the average response time, tt is the think time and tp is the measured throughput.
Setting tt to 60 seconds, rt to 0.85 seconds and tp to 119.44 queries/sec (430,000 queries/hour), the above formula shows that the T4-4 server will support 7,300 concurrent users with a think time of 60 seconds and an average response time of 0.85 seconds.
For more information see chapter 3 from the book "Quantitative System Performance" cited below.
Quantitative System Performance
Computer System Analysis Using Queueing Network Models
Edward D. Lazowska, John Zahorjan, G. Scott Graham, Kenneth C. Sevcik
Oracle Database 11g – Oracle OLAP
- SPARC T4-4 Server
Oracle Database 11g Release 2
Copyright 2012, Oracle and/or its affiliates. All rights reserved. Oracle and Java are registered trademarks of Oracle and/or its affiliates. Other names may be trademarks of their respective owners. Results as of 11/2/2012.