07.02.2013 Views

Best Practices for SAP BI using DB2 9 for z/OS - IBM Redbooks

Best Practices for SAP BI using DB2 9 for z/OS - IBM Redbooks

Best Practices for SAP BI using DB2 9 for z/OS - IBM Redbooks

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Objectives<br />

The goal <strong>for</strong> the multi-user query scenarios is to measure the following KPIs <strong>for</strong><br />

5 TB on 27 blades, 15 TB on 81 blades, and 25 TB on 135 blades,<br />

► Query throughput rate in terms of business reports per hour and InfoCube<br />

queries per hour<br />

► Average number of records touched per query in the <strong>BI</strong> Accelerator<br />

► Average report response time<br />

► CPU consumption on blades and <strong>BI</strong> server<br />

Description<br />

When a query is issued by a user, <strong>SAP</strong> <strong>BI</strong> on System p checks whether the data<br />

to be queried has been <strong>BI</strong> Accelerator indexed. If it has, query processing is<br />

directed to <strong>SAP</strong> <strong>BI</strong> Accelerator. Then if the data is not in <strong>BI</strong> Accelerator blade<br />

memory, it must be read from GPFS into blade memory. In preparation <strong>for</strong> query<br />

processing, all cube data can be pre-loaded from GPFS into blade memory<br />

during index creation or off-shift hours. This speeds up the query processing by<br />

eliminating I/O during production.<br />

For our query measurements, all <strong>BI</strong> Accelerator indexes were pre-loaded into<br />

memory.<br />

Of all the records in the InfoCubes, only selected rows (QDBSEL) are relevant to<br />

the query. As with RDBMS, filtering of all records in the InfoCube is done to<br />

identify the ones relevant to the query <strong>for</strong> further processing. Typically, indexes<br />

are used to do this. In this case, <strong>BI</strong> Accelerator column-based processing and<br />

compression have a clear advantage over an RDBMS in identifying and<br />

processing the relevant records. The selected records are summarized or<br />

aggregated concurrently by a number of blades, then the records are transferred<br />

to an OLAP engine in an <strong>SAP</strong> <strong>BI</strong> appserver residing on System p. The number of<br />

records transferred is referred to as QDBTRANS. At the <strong>SAP</strong> <strong>BI</strong> appServer,<br />

server rendering is per<strong>for</strong>med be<strong>for</strong>e sending the query report to users.<br />

Note that the queries that we designed <strong>for</strong> this study had high selectivity, with an<br />

average QDBSEL of 26–46 million records per query, <strong>for</strong> our 25 TB tests. Still,<br />

the <strong>BI</strong> Accelerator does tremendously well with an average response time of less<br />

than 20 seconds.<br />

Measurement results<br />

One benefit to <strong>using</strong> <strong>BI</strong> Accelerator is that no explicit tuning <strong>for</strong> the cubes is<br />

required to prepare <strong>for</strong> query processing. So there are no aggregates to be<br />

defined and no additional secondary indexes are needed.<br />

Chapter 11. Project Jupiter: large <strong>BI</strong> Accelerator scalability evaluation 247

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!