19.06.2013 Views

DB2 UDB for z/OS Version 8 Performance Topics - IBM Redbooks

DB2 UDB for z/OS Version 8 Performance Topics - IBM Redbooks

DB2 UDB for z/OS Version 8 Performance Topics - IBM Redbooks

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

4.5.2 Conclusion<br />

The per<strong>for</strong>mance improvements in a data sharing environment are not as pronounced, with<br />

only a 19% improvement in DBM1 CPU used, compared with 29% in non-data sharing. The<br />

main reason is a higher fixed cost with data sharing.<br />

Remember that our measurements were taken using an I/O intensive workload. You might not<br />

see as dramatic an improvement in CPU savings as we did. However our test cases show a<br />

CPU time reduction of up to 8% <strong>for</strong> I/O intensive transaction workload, and as much as 10%<br />

<strong>for</strong> more I/O intensive batch types of workloads.<br />

Although these are still significant CPU savings to be obtained in a data sharing environment,<br />

the CPU savings are generally not as great as in a non-data sharing environment. This is<br />

primarily due to a higher fixed cost overhead with data sharing.<br />

4.5.3 Recommendations<br />

4.6 IRLM V2.2<br />

Long term page fixing <strong>DB2</strong> buffer pools is effectively a trade-off between real storage and<br />

per<strong>for</strong>mance. It is available in V8 CM and beyond.<br />

Long term page fixing of <strong>DB2</strong> buffer pools typically results in a CPU savings in the range of<br />

0% to as great as 10%. The per<strong>for</strong>mance benefits are inversely proportional to the buffer pool<br />

hit ratio. The higher the hit ratio, the lower the potential per<strong>for</strong>mance benefit. We there<strong>for</strong>e<br />

recommend you first page fix your smaller buffer pools with a low hit ratio and frequent I/O,<br />

(<strong>for</strong> example sequential workloads). You then should consider page fixing your other buffer<br />

pools, buffer pool by buffer pool, to minimize potential impact on other concurrently running<br />

workloads. However, you need to be a little more cautious when considering larger buffer<br />

pools. Page fixing buffer pools may drive system paging rates higher, thereby impacting<br />

overall system per<strong>for</strong>mance.<br />

We cannot overstress the importance of having sufficient real storage to back up your buffer<br />

pool 100%. Having 99.9% is not good enough, as the LRU buffer steal algorithm introduces<br />

paging if there is insufficient real storage to back up your buffer pool completely. There<strong>for</strong>e,<br />

make sure your buffer pool is fully backed by real storage be<strong>for</strong>e you start using PGFIX(YES).<br />

<strong>DB2</strong> V8 is shipped with IRLM V2.2, which is now a 64-bit application. With IRLM V2.2, <strong>DB2</strong><br />

locks always reside above the 2 GB bar.<br />

IRLM V2.2 ships with both a 31-bit version of the modules, as well as a 64-bit capable<br />

version. If the operating system is able to handle 64-bit, the 64-bit version is loaded. However,<br />

the interface to IRLM V2.2 is still 31-bit. The 64-bit version requires the z/Architecture as well<br />

as z/<strong>OS</strong> 1.3. <strong>DB2</strong> V8 requires IRLM V2.2 in 64-bit mode.<br />

IRLM V2.2 is also shipped with IMS V9. However, note that at the time of writing this book,<br />

IRLM V2.2 does not support earlier versions of <strong>DB2</strong> or IMS.<br />

IRLM V2.2 supports a far greater number of locks than V2.1. IRLM V2.1 which is shipped with<br />

<strong>DB2</strong> V7, is a 31-bit application and is there<strong>for</strong>e restricted to the 2 GB address space limit. The<br />

maximum number of concurrently held locks in IRLM V2.1 is approximately 8 million locks, (2<br />

GB/approximately 260 bytes per lock). IRLM V2.2 which is shipped with <strong>DB2</strong> V8, is a 64-bit<br />

application and there<strong>for</strong>e not restricted to the 2 GB address space limit. As the average lock<br />

size in <strong>DB2</strong> V8 is increased to approximately 540 bytes, only 3.7 million locks can be<br />

Chapter 4. <strong>DB2</strong> subsystem per<strong>for</strong>mance 171

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

Saved successfully!

Ooh no, something went wrong!