23.10.2014 Views

Advanced POWER Virtualization on IBM System p5 - Previous ...

Advanced POWER Virtualization on IBM System p5 - Previous ...

Advanced POWER Virtualization on IBM System p5 - Previous ...

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.

►<br />

►<br />

►<br />

If you dynamically remove a virtual processor, you cannot specify an<br />

identificati<strong>on</strong> for particular virtual processor to be removed. The operating<br />

system will choose the virtual processor to be removed.<br />

Shared processors may render AIX 5L affinity management useless. AIX 5L<br />

will c<strong>on</strong>tinue to utilize affinity domain informati<strong>on</strong> as provided by firmware to<br />

build associati<strong>on</strong>s of virtual processors to memory, and it will c<strong>on</strong>tinue to<br />

show preference to redispatching a thread to the virtual processor that it last<br />

ran <strong>on</strong>.<br />

An uncapped partiti<strong>on</strong> with a weight of 0 has the same performance impact as<br />

a partiti<strong>on</strong> that is capped. The HMC can dynamically change either the weight<br />

or a partiti<strong>on</strong> from capped to uncapped.<br />

Dispatching of virtual processors<br />

There is additi<strong>on</strong>al processing associated with the maintenance of <strong>on</strong>line virtual<br />

processors, so you should carefully c<strong>on</strong>sider their capacity requirements before<br />

choosing values for these attributes.<br />

Under AIX 5L V5.3 ML3, a new feature is introduced to help manage idle virtual<br />

processors.<br />

Virtual processors have dispatch latency since they are scheduled. When a<br />

virtual processor is made runnable, it is placed <strong>on</strong> a run queue by the <str<strong>on</strong>g>POWER</str<strong>on</strong>g><br />

Hypervisor, where it waits until it is dispatched. The time between these two<br />

events is referred to as dispatch latency.<br />

The dispatch latency of a virtual processor depends <strong>on</strong> the partiti<strong>on</strong> entitlement<br />

and the number of virtual processors that are <strong>on</strong>line in the partiti<strong>on</strong>. The capacity<br />

entitlement is equally divided am<strong>on</strong>g these <strong>on</strong>line virtual processors, so the<br />

number of <strong>on</strong>line virtual processors impacts the length of each virtual processor's<br />

dispatch. The smaller the dispatch cycle, the greater the dispatch latency.<br />

At the time of writing, the worst case virtual processor dispatch latency is 18<br />

millisec<strong>on</strong>ds, since the minimum dispatch cycle that is supported at the virtual<br />

processor level is <strong>on</strong>e millisec<strong>on</strong>d. This latency is based <strong>on</strong> the minimum<br />

partiti<strong>on</strong> entitlement of 1/10 of a physical processor and the 10 millisec<strong>on</strong>d<br />

rotati<strong>on</strong> period of the Hypervisor's dispatch wheel. It can be easily visualized by<br />

imagining that a virtual processor is scheduled in the first and last porti<strong>on</strong>s of two<br />

10 millisec<strong>on</strong>d intervals. In general, if these latencies are too great, then clients<br />

may increase entitlement, minimize the number of <strong>on</strong>line virtual processors<br />

without reducing entitlement, or use dedicated processor partiti<strong>on</strong>s.<br />

Chapter 2. <str<strong>on</strong>g>Virtualizati<strong>on</strong></str<strong>on</strong>g> technologies <strong>on</strong> <strong>System</strong> p servers 43

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

Saved successfully!

Ooh no, something went wrong!