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

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

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

Table 4-11 IRLM CPU - V2.1 vs. V2.2<br />

<strong>DB2</strong> V7 V8 V8 Delta <strong>for</strong> P1 %<br />

(V8-V7) / V7<br />

IRLM (PC=YES) 2.1 2.2 2.2<br />

Lock Protocol n/a 1 2<br />

Group ITR<br />

(commits/sec)<br />

IRLM CPU<br />

msec/commit<br />

We can make some judgements here to assess the impact on CPU the IRLM V2.2 has over<br />

the IRLM V2.1. Keep in mind that IRLM V2.2 is now a 64-bit application and the locks are<br />

managed above the 2 GB bar. These figures show only a modest 1.8% increase in CPU with<br />

this workload. The IRLM is generally not a significant user of CPU compared with other <strong>DB2</strong><br />

address spaces such as the DBM1 address space. So, this slight increase is not really<br />

considered significant.<br />

Significant savings in the IRLM CPU can be seen in the third test, when Locking Protocol<br />

Level 2 was used, compared with the first test under <strong>DB2</strong> V8. This CPU savings can directly<br />

be attributed to the Locking Protocol Level 2, which significantly reduces XES contention<br />

resolution. Refer to Chapter 8, “Data sharing enhancements” on page 319 <strong>for</strong> a more detailed<br />

discussion of the new locking protocol implemented in <strong>DB2</strong> V8.<br />

PC=NO vs. PC=YES<br />

In the past, the IRLM PC parameter, an option on the IRLM startup procedure JCL,<br />

determined where the IRLM maintains its locks. PC=NO instructs the IRLM to maintain its<br />

lock structures in ECSA, which is governed by the MAXCSA parameter. PC=YES instructs<br />

the IRLM to maintain its lock structures in the IRLM private region. PC=YES causes<br />

cross-memory linkage PC constructs to be used between <strong>DB2</strong> and the IRLM, rather than the<br />

cheaper branch constructs.<br />

In the past, PC=NO was the general recommendation <strong>for</strong> per<strong>for</strong>mance, and PC=NO is the<br />

default in <strong>DB2</strong> V7. However, Cross Memory calls have become less expensive and recent<br />

enhancements in lock avoidance techniques have greatly reduced the advantage of using<br />

PC=NO.<br />

Prior to <strong>DB2</strong> V8, PC=YES has been the recommendation <strong>for</strong> high availability. PC=YES helps<br />

you avoid “out of storage” problems related to exhausting ECSA. It also allows you to<br />

decrease the ECSA size, and thus increase the private region size. It also allows you to raise<br />

lock escalation limits and max locks per user limit in DSNZPARM, reducing the frequency of<br />

lock escalations.<br />

For some workloads, going from PC=NO to PC=YES may increase the CPU cost of each<br />

IRLM request by as much as 10%. This is because the number of instructions required to<br />

process a lock increase by about 5%, and cross memory instructions are generally more<br />

expensive to process than normal instructions. However if 2% of the overall CPU time is<br />

consumed by the IRLM, the overall impact of having PC=YES is 0.2%, (10% * 2%), which is<br />

probably not noticeable. So, even in <strong>DB2</strong> V7, the availability benefits of moving from PC=NO<br />

to PC=YES outweigh the very slight per<strong>for</strong>mance benefits of PC=NO in most cases.<br />

In any event, the PC and MAXCSA parameters are no longer used in the IRLM V2.2.<br />

However, you must still specify them in the IRLM JCL <strong>for</strong> compatibility reasons.<br />

Delta <strong>for</strong> P2 %<br />

(V8-V7) / V7<br />

823.56 804.69 933.41 -2.29% +13.34%<br />

0.330 0.371 0.335 0.379 0.014 0.014 +1.85% -96.01%<br />

Chapter 4. <strong>DB2</strong> subsystem per<strong>for</strong>mance 173

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

Saved successfully!

Ooh no, something went wrong!