Re: Breaking up one partition into two.

To take Larry's comments one small step further, lately I have been turning off automatic performance adjustment for customers that have the following profile:
1) Someone capable of monitoring the box periodically during the day (operator or system admin)
2) SQL being used in the applications and/or from ODBC/OLE connections
3) Extensive use of ILE with named activation groups

The reasons are simple. If the SQE sees changes in memory configuration it will re-optimize queries. It has to since the last plan would not work when the available memory is different. While this is a really fast process, I have seen significant throughput changes particularly with SQL heavy products such as Web Query and the like. If you have a web site running, regardless of the type, it's almost certainly using SQL. If the system is memory or CPU constrained the positive effect is larger. Downside, it requires someone to manage Night/Day etc. (Back to the future with those CL pgms)

With ILE, I have run into situations were the programs can behave in unexpected ways and have very poor performance when customers that use ILE to its fullest. Moving these processes into a unique memory pool (sometimes even a private pool) seems to mitigate those issues.

With products that use the IFS extensively (Domino, WebSphere, Zend Server, Apache, etc) I put those into a separate pool because the I/O characteristics and the paging patterns are really different than more traditional applications.

The real message both Larry and I are trying to convey is simple. Give IBM i its due and provide it enough memory in *MACHINE and *BASE memory pools. Move most, if not all user jobs out of *BASE. Tune from there. Your system will like you for it.

Jim Oberholtzer
Chief Technical Architect
Agile Technology Architects


On 7/23/2011 8:51 AM, DrFranken wrote:

To answer just the pool question:

Jim is right that 5400 is a very big number and if that's for an
entire 5 or more minute time frame that needs addressing! Second your
comment about numbers from a B10 being applicable to a POWER6 machine is
also accurate. A B10 would have already died before it sustained 5400
faults a second for any amount of time.

The key to pools is that different work is kept apart. Domino has
various processes that run, something like 20 jobs per sever as I recall
and most if not all of them are threaded meaning the actual number of
operating tasks is quite large. Each thread needs some memory and an
activity level in order to run. NOT every thread is running at all
times. By having Domino in its own pool there isn't any conflicting
demand for memory resource there and Domino tasks don't get paged out
(at least most of them won't.) Thus when they are needed they simply run
without fanfare or additional system effort to bring them back into memory.

Meanwhile back in *BASE (or better yet your batch run pool) that AP
check run (which we consultants are very fond of by the way:-) ) gets
submitted. It also needs memory resources. If it uses SQL the query
optimizer looks around to see how much memory is available and uses
whatever it can to do its work. If there are jobs there in a wait state
those jobs could get paged out for the AP job. Of course the jobs
finishes and gives up it's memory to the next batch job. This is good,
as you have a completely different type of task (Start, run hard, Done)
in this pool -vs- the Domino subsystem (Start, run as needed forever,)
So even if the AP job has a much higher run priority (remember I'm good
with that!) It just gets done faster it doesn't shove Domino out of
memory. And since you have lots of CPUs it may not make any difference
to Domino performance at all.

That's why different memory pools for types of tasks is a good
idea. It's the same reason you don't run you interactive jobs in *BASE.

- Larry "DrFranken" Bolhuis

On 7/22/2011 4:22 PM,rob@xxxxxxxxx wrote:

> We do already have separate IP addresses for each domino server and a
> separate one for i, (and another separate one with it's own card for
> Mimix).
>
> Back in the day we used to manage memory ourselves. We used to have two
> job schedule entries NIGHT and DAY that would shift memory between batch
> and interactive. We kind of like this theory of autonomics and having the
> system fix itself.
>
> Our concern with creating multiple pools is that you end up allocating
> resources that are not available to other processes. For example, if I
> bust *BASE into *BASE and DOMINO. I would allocate x memory to *BASE and
> y memory to DOMINO. If one is not using it and the other is not (and
> QPFRADJ is turned off) then I end up with a bunch of inventory not being
> used, when other processes would love to have it.
>
> I think one theory is that by busting out DOMINO from *BASE you can
> control a spike in either from affecting the other. Is there some other
> reason to do so? Like, "paging option"? I heard someone mention run
> priority. I thought that was something at the job level. Upon further
> review I do see that subsystems support "classes". From there you can
> control run priority, time slices, threads, etc. I guess I just don't
> know enough "why" as to why I would want to do this other work. Like
> restricting down threads, etc.
>
> In general "I thought" our performance was good. Throwing money at a
> problem greatly simplifies management. There's a lot that I used to do on
> AS/400's that I don't mess with so much anymore on Power equipment. Also,
> our people thrive on Domino. If I throttled down a hundred Domino users
> so one persons A/P check run would have priority it would not go over
> well.
>
> I have forgotton so much, like why faulting is bad, but even more so, the
> guidelines as to what levels constitute high faulting. I can't believe
> the number one used on a B10 is applicable to a 9117-MMA.
>
> I am seeing that a separate partition for this is probably a bad idea. But
> I am still trying the grasp why a separate pool might help.
>
>
> Rob Berendt