Mainframe Capacity planning.

[Ajay Pandit]

Hi,

MIPS is usually used for LPAR level usage of a mainframe. And there is another thing called as SU, which is used to measure the amount service done by the infrastructure. But if I need to do a Capacity Planning for the underlying infrastructure of mainframes where can I start with? Can any of these (MSU, SU) be of some help to plan for that?

[Robert Sample]

Converting SU to MSU is real simple -- divide SU by 1 million and you'll get MSU (Millions of Service Units). Converting MSU (which is used much more than SU) to MIPS (and vice versa) can be done but it is not a simple conversion. IBM does not support the use of MIPS (or MSU for that matter) for capacity planning or performance management, although MIPS is still used by many sites. You can understand this by thinking about a site with a box having 16 processors -- some CP, some zIIP, some zAAP -- so the concept of "million instructions per second" doesn't consider how many processors are being used, nor how they are being used.

Capacity planning and performance management can be quite complex due to the potential questions involved -- is the processor capped? are there specialty engines used? how many LPARs are defined? how are the processors split between the LPARs? what weights are given the various LPARs? Furthermore, you need to consider not just TCB but also SRB and I/O time when looking at capacity and performance.

IBM has established a number of different benchmark workloads (batch, online, low I/O, medium I/O, high I/O, etc) that can be used to help with capacity planning. The site matches their workload characteristics against the different IBM workloads to find the best match, then uses that workload for their planning. Watson and Walker has some good material on capacity planning, and if your site has MXG installed then reading through some of the members of the MXG source PDS can be extremely helpful. Otherwise, you'll need to use Google and do some research.

[Ajay Pandit]

Thanks Robert. I'm reading manuals to understand all these one by one. But would you like to recommend me to read a specific book ot ttorial so I that I can have better understading of what you have summed up here:

Capacity planning and performance management can be quite complex due to the potential questions involved -- is the processor capped? are there specialty engines used? how many LPARs are defined? how are the processors split between the LPARs? what weights are given the various LPARs? Furthermore, you need to consider not just TCB but also SRB and I/O time when looking at capacity and performance.

[Robert Sample]

Start with the IBM Redbook ABC's of z/OS System Programming Volume 11 http://www.redbooks.ibm.com/redbooks/pdfs/sg246327.pdf which covers Capacity Planning, Performance Management, RMF, and SMF. Also visit the CMG (Computer Measurement Group) web site at http://www.cmg.org as they have lots of good information.

[Gerhard_Adam]

MSU's should never be used for capacity planning. SU's do NOT convert to MSU's because they represent two entirely different metrics. SU's are "charged" to specific units of work so that they represents a normalizing value to determine CPU or I/O usage. MSU's are defined as a capacity metric for software licensing purposes and do not reflect the actual service delivery capability of the machine.

[rohithkm]

I have been tasked to pull the mips report for every 15mins a day on Lpar basis. There is no rmf, sas/mxg. Took help of CA Explore Report writer product to get to smf record 72 and get the CPU Service Unit and Total Service Unit in the report. Now looking for a calculation which can yield the Mips or Msu usage.

[ Post made via Android ]

[Robert Sample]

Type 72 records are not what you use for this - the type 70 records are what you want. I will post more tomorrow on the calculations based on the type 70 records.

[ Post made via Android ]

[rohithkm]

Thanks Robert. I get the below options from smf type 70 and 72 from ca explore report writer.

[ Post made via Android ]

[Robert Sample]

The reason you do not want the type 72 records is because they are Workload Activity (WLM) records and hence if you have report classes defined these will duplicate the service classes -- it is very easy to double-count CPU time when using type 72 records. A given WLM workload may have a service class defined, or a report class defined, or both a service class and a report class defined. So you have to know which report classes need to be counted and which do not. I'm assuming you have verified that your site generated SMF interval data for quarter hours (not all sites do -- some use half hour intervals). Note that the type 72 records are not by LPAR -- but they do include the partition number of the partition that wrote the record so you can translate the partition number to the LPAR name.

The type 70 records are based on LPAR and CPU. A system z CPU can be a CP (what most people consider to be the computer), an ICF (coupling facility), an IFL (Linux processor), a SAP (system assistance processor -- used for I/O), a zAAP (no longer used but still part of older machines and used for Java), or a zIIP (integrated information processor -- used for DB2 and certain other types of workloads, and for Java on newer machines). The ICF, IFL, SAP processors can be ignored if present in the type 70 records. Depending upon the machine your site runs, there may be an adjustment factor required for zIIP / zAAP CPU time (since they run at full speed no matter what speed your site's machine runs at). I'm not familiar with CA Explore so I don't know how (or if) it aggregates quarter hour records. The z14 can have up to 170 processors in various combinations of CP, ICF, IFL, SAP, zIIP -- so what, exactly, is the MIPS measuring in such an environment? How many processors of the various types are installed at your site? And an LPAR can have a dedicated processor or it can share a processor, so calculating MIPS (MSU) can be challenging.

I have been tasked to pull the mips report for every 15mins a day on Lpar basis.

I'm assuming you're planning on preparing this report weekly (or daily, depending upon how often your SMF records are collected) since doing this during the day would be a big waste of resources. When doing the calculations, you can either use the LPAR values or the service units (which have to be multiplied by the WLM factors which you may have to get from WLM); if you use service units (after accumulating the products, you divide by 1,000,000 to convert SU to MSU) you may need to use the LSPR ITR table values for your machine to convert MSU to MIPS.

Using the type 70 records gives you CPU seconds, but the calculations can get arduous since they are based on each LPAR. Compute CPU time as interval time (SMF70INT in the Product section) minus CPU wait time (SMF70WAT in the CPU Data section) and accumulate for the CP processors; if the processor is a zIIP (or zAAP) you multiply the CPU time by normalization factor (SMF70NRM in the CPU Control section) and divide by 256 to get CP-equivalent CPU time. Accumulate for the processors for that time period. Note that shared processors may be double counted by this method since I'm not sure how their CPU time gets reported.

[Robert Sample]

Here https://share.confex.com/share/117/webp ... essors.pdf is a Share presentation on the type 70 and 72 records. It is worth reviewing.

[rohithkm]

Hi Robert,
Good day.

Thanks a lot for the information. I went through the share presentation and to understand different scenarios you mentioned.

The daily SMF GDG is split into various departments and one for Performance. Then I pass the performance gdg that has smf type 70-79 to CA Explore Report Writer and it generates the output as mentioned in the variable class column with respect corresponding to the Data column which we want it to generate from previous screenshots. And it provides filter option to generate the report for every 15mins and other intervals too.

Have checked that we use only 2GPs and no special processors and 3Lpars shares the processors. We pass the performance gdg as input on lpar specific to generate the reports. And to generate the report at the end of the day after the performance gdg is generated that has entire day statistics from 00:00 to 23.15 on lpar basis.

You mentioned to use Lpar values and then found this option in the ca report writer which uses type70 to generate logical CPU Processor Usage in % and Actime time with SMF70PDT data. Have attached this screenshot.

I am sure now we have to go with SMF type70 for Mips or MSU stats to be accurate in terms of lpar specific.

And checking other possibility with BMC CMF Analyzer product that has replaced RMF at our site and it collects Smftype 70-79 records and dumps into smf man datasets and this is used to get the smf daily gdg that has dump of all man files at eod and splits into various outputs from which I take performance as input for report generation.

Thanks Again Robert for the information. Please let me know your thoughts on my above points and will further research and come back with the updates. Attached the screenshot with logical processor usage stats from ca report writer using smftype70.

[ Post made via Android ]

[nicc]

locked. Old topics should not be revived (but they can be referred to).