In short, giant LPARs can definitely be problematic. Similarly, too small LPARs 
can be problematic. Somewhere in the middle is ideal, but where that is will 
depend. 

First off, the most significant impact is you don't want LPARs whose processor 
count is so high that it crosses drawers. (In most cases this count would be 
zIIP + GP.) So that's a high water mark you definitely want to avoid that's 
around the low 30s for relatively recent machines. Similarly you don't want 
LPARs with so much memory that they cross drawers, although that may be 
slightly less problematic. In both cases, if you have CPs referencing memory or 
L4 cache on another drawer, access times will be longer. How big the impact 
will be depends on how much of that cross-drawer access is going on.

Secondly, there's internal housekeeping and locks and latches that are impacted 
by having more CPs. How much this impacts things depends again on the workload. 

In contrast, more LPARs does add more MSU consumption for certain system tasks 
like monitors and the like. I.E. add an extra LPAR and it's an extra RMF/CMF 
that's running collecting/writing data. And extra whatever other monitors you 
have. And while adding a third system to the sysplex adds nowhere near the 
overhead that adding a second adds, there still can be some. E.G. resolving 
lock contention now involves potentially 3 systems instead of 2. So I wouldn't 
look to add systems willy-nilly where I have a sysplex of lots of small LPARs, 
all that are part of a single data sharing group. (I'm less concerned about 
adding dev/test LPARs that have minimal sharing with the production LPARs 
already part of the sysplex.)

My general rule of thumb, is that: <10 busy CPs I'm not concerned; more than 20 
busy CPs I might start to wonder (not necessarily worry) a bit; more than a 
drawer, I definitely want to have a discussion about it. I'm not sure that 
those are perfect thresholds, and I'd generally temper that with specific 
details about the situation. 

Somebody mentioned not having enough CPs such that production LPARs can't get 
any high-polarity CPs. I agree that's a growing problem that we're seeing with 
multiple customers due to the wider steps between the sub-capacity settings on 
the large machines. In general, going from zero to multiple high-polarity 
processors for important LPARs is important, but tweaking weights to convert a 
medium to a high on an LPAR that already has multiple highs is probably not 
going to be significant. (That's the general conclusion of my GSE presentation 
this Thursday.)

----------------------------------------------------------------------
For IBM-MAIN subscribe / signoff / archive access instructions,
send email to [email protected] with the message: INFO IBM-MAIN

Reply via email to