By Brian-Oracle on Mar 26, 2013
Oracle's SPARC T5-8 server delivered world record query performance with near real-time analytic capability using the Oracle OLAP Perf Version 3 workload running Oracle Database 11g Release 2 on Oracle Solaris 11.
The maximum query throughput on the SPARC T5-8 server is 1.6x higher than that of the 8-chip Intel Xeon E7-8870 server. Both systems had sub-second response time.
The SPARC T5-8 server with the Oracle Database demonstrated the ability to support at least 600 concurrent users querying OLAP cubes (with no think time), processing 2.93 million analytic queries per hour with an average response time of 0.66 seconds per query. This performance was enabled by keeping the entire cube in-memory utilizing the 4 TB of memory on the SPARC T5-8 server.
Assuming a 60 second think time between query requests, the SPARC T5-8 server can support approximately 49,450 concurrent users with the same 0.66 sec response time.
The SPARC T5-8 server delivered 4.3x times the maximum query throughput of a SPARC T4-4 server.
The workload uses a set of realistic BI queries that run against an OLAP cube based on a 4 billion row fact table of sales data. The 4 billion rows are partitioned by month spanning 10 years.
The combination of the Oracle Database with the Oracle OLAP option running on a SPARC T5-8 server supports live data updates occurring concurrently with minimally impacted user query executions.
Oracle OLAP Perf Version 3 Benchmark
Oracle cube base on 4 billion fact table rows
10 years of data partitioned by month
|0 sec think time||60 sec think time|
|8-chip Intel Xeon E7-8870||1,823,000||120||30,500||0.19|
Configuration Summary and Results
SPARC T5-8 Hardware Configuration:
4 TB memory
Oracle Database 11g Release 2 (184.108.40.206) with Oracle OLAP option
Sun Server X2-8 Hardware Configuration:
512 GB memory
Oracle Database 11g Release 2 (220.127.116.11) with Oracle OLAP option
SPARC T4-4 Hardware Configuration:
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 3 benchmark is a workload designed to demonstrate and stress the ability of the OLAP Option to deliver fast query, near real-time updates and rich calculations using a multi-dimensional model in the context of the Oracle data warehousing.
The bulk of the benchmark entails running a number of concurrent users, each issuing typical multidimensional queries against an Oracle 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.
While the core of every OLAP Perf benchmark is real world query performance, the benchmark itself offers numerous execution options such as varying data set sizes, number of users, numbers of queries for any given user and cube update frequency. Version 3 of the benchmark is executed with a much larger number of query streams than previous versions and used a cube designed for near real-time updates. The results produced by version 3 of the benchmark are not directly comparable to results produced by previous versions of the benchmark.
The near real-time update capability is implemented along the following lines. A large Oracle cube, H, is built from a 4 billion row star schema, containing data up until the end of last business day. A second small cube, D, is then created which will contain all of today's new data coming in from outside the world. It will be updated every L minutes with the data coming in within the last L minutes. A third cube, R, joins cubes H and D for reporting purposes much like a view might join data from two tables. Calculations are installed into cube R. The use of a reporting cube which draws data from different storage cubes is a common practice.
Query users are never locked out of query operations while new data is added to the update cube. The point of the demonstration is to show that an Oracle OLAP system can be designed which results in data being no more than L minutes out of date, where L may be as low as just a few minutes. This is what is meant by near real-time analytics.
Key Points and Best Practices
Update performance of the D cube was optimized by running update processes in the FX class with a priority greater than 0. The maximum lag time between updates to the source fact table and data availability to query users (what was referred to as L in the benchmark description) was less than 3 minutes for the benchmark environment on the SPARC T5-8 server.
Building and querying cubes with the Oracle OLAP option requires a large temporary tablespace. Normally temporary tablespaces would reside on disk storage. However, because the SPARC T5-8 server used in this benchmark had 4 TB of main memory, it was possible to use main memory for the OLAP temporary tablespace. This was done by using files in /tmp for the temporary tablespace datafiles.
Since typical BI users are often likely to issue similar queries, either with the same, or different, constants in the where clauses, setting the init.ora parameter "cursor_sharing" to "force" provides for additional query throughput and a larger number of potential users.
Assuming the normal Oracle initialization parameters (e.g. SGA, PGA, processes etc.) are appropriately set, out of the box performance for the OLAP Perf workload should be close to what is reported here. Additional performance resulted from (a)using memory for the OLAP temporary tablespace (b)setting "cursor_sharing" to force.
For a given number of query users with zero think time, the main measured metrics are the average query response time and the query throughput. A derived metric is the maximum number of users the system can support, with the same response time, assuming some non-zero think time. The calculation of this maximum is 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 0.66 seconds and tp to 815 queries/sec (2,934,000 queries/hour), the above formula shows that the SPARC T5-8 server will support 49,450 concurrent users with a think time of 60 seconds and an average response time of 0.66 seconds.
For more information about the "response-time law" see chapter 3 from the book "Quantitative System Performance" cited below.
- SPARC T5-8 Server
- SPARC T4-4 Server
Oracle Database 11g – Oracle OLAP
my.oracle.com oracle.com OTN
Oracle Database 11g Release 2
Quantitative System Performance
Computer System Analysis Using Queueing Network Models
Edward D. Lazowska, John Zahorjan, G. Scott Graham, Kenneth C. Sevcik
Lazowska UW site
Copyright 2013, 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 03/26/2013.