It’s also important to look at VP-to-entitlement ratios. Ideally the ratio should be 2.5 or less. Anything above 4.0 is performance unfriendly, especially on multi-node systems (770 and above). When the VP is dispatched but its home node is busy, the system uses scheduler resource affinity domains (SRADs) at run time to determine the best core to dispatch it to. This core could be local, near or far. However, the memory might still be allocated on DIMMs attached to the home or some other core. As the LPARs get busy, this is more and more likely to happen. If memory is local, the bandwidth on the POWER7 is 68GB/s per memory controller. If it’s near, it goes to 40-50GB/s, and far memory is about 23-26GB/s.
Clearly, you take a performance hit if memory isn’t on the DIMMs attached to the home core. On a very busy system with lots of LPARs, it can lead to thrashing of the memory subsystem as well as affinity issues. This is more likely to happen with lots of VPs and low entitlement in the LPAR because the LPAR will be spread further in raw throughput mode. The lssrad command can show if this is happening. You can also run the Hardware Management Console (HMC) scanner to get information on servers and calculate the VP-to-entitlement ratios, keeping them within the values recommended by IBM’s Nigel Griffiths
after we have added cores to our prod system we started to see severe fardispatching
now how can we check our system under this guidance at the document?