01 August 2012

Getting the CPU-Measurement Facility to work in z/VSE

At long last - the promised guest blog entry from Martin Trübner! Vielen dank!

Let me first introduce myself



My name is Martin Trübner. I am German and work from here, but did (when being on site was vital) gigs all over Europe and the US. I was educated and did my first job in z/OS predecessors but did my first sysgen on DOS/VS R31. During my activities I do run into z/OS and it is a wonderful environment. You can do everything and everything can be avoided and/or forbidden (through real measures), but not so in DOS/VS. Today I am bit-wise active in z/VSE and CICS/TS, and as an application person (with enough background to be dangerous) in z/OS. 







First Encounter



With the announcement of the z196 I stumbled over the CPU-Measurement Facility (CPU-MF). I was delighted to read that it is available on the z10 as well. The various counters did not really interest me, but the sampling was something for which I could imagine a very rewarding use case. If you want details of what is collected, here is a picture with mouseover explanations. But I read that it is only available for native operating systems. Since most of the z/VSE shops do run with z/VM on top, there was no point for me to investigate further. 

During a presentation at the end of June 2012, I was alerted by a single sentence that read: 
New z/VM CPU MF COUNTERS support – APAR VM64961
I thought to myself: “Now this is good news. Let's dive into the details”. 



Preliminary checks

I Googled for it and found a few presentations, the book that documents the instructions and structures used for it, and some bits and pieces about it; I must also not forget documentation regarding Hardware Instrumentation Services (HIS) for z/OS (thanks to "E."). Here are my findings:
  • For sampling, it needs a setting in the SE/HWC
  • For VM it needs the above PTF
  • z/OS has HIS that explores it and makes lots of extra information  available
  • z/HISR from Phoenix has software that uses the z/OS HIS data.
  • The CPU-MF "POP-equivalent" manual, Load-Program-Parameter and the Set-Program-Parameter and CPU-Measurement Facilities (IBM manual SA23-2260-02) says that this facility is a prerequisite for the CPU-measurement sampling facility.
  • The structures for sampling are very clever and suitable for any number of samples
  • There is even provision for emptying buffers while sampling is active  
  • There is even extra space for extension (as it was when TRAP–ping was introduced, shortly before 64-bit was introduced).

Setup in z/VSE / Problems

Nothing is a show-stopper, so I decided to go for it and write a POC with only 1–minute sampling at every 2 milliseconds (just one SDB (sampling data buffer) at 1MB), and here is what happened…

(Please do not forget- this is all occuring in z/VSE 5.1, which is the first version of VSE that supports 64-bits virtual.)
  • My hardware does not have it. Okay, there are customers with a z10/z196/z114 that willing to help but the switching on via the SE/HWC is not an everyday activity. So here we have a delay.
  • It needs IARV64 – my first chance to use it – one more reason to go for it.
  • For debugging I would like to display what I “prepared for the HW”/”got back from HW”. 
  • Ah…well, SDUMPX does support areas above the bar–but how do I get an address space STOKEN. I know what it is but heck – should I really go for control blocks to get it? Thanks to "I.", who showed me an undocumented operand of ALESERV. So now we can display these areas even without the famous 0C3 or my debugger (TRAPPER, which does understand 64-bit storage).(Ray adds: I know of this operand. It can also be used to access other partitions via access registers.)
  • IARV64 REQUEST=PAGEFIX,RANGLIST=RLIST_PTR,NUMRANGE=1 abends with a ABEND 2C5 reason 0598 which reads: “0598 IARV64: Parameter list or range list must be in 31-bit storage”. How can I guarantee 31 bit storage? I can do 24, below the bar, and any boundary- but 31? With the help of "I." it was identified as typo (meaning “not 64”) and a coding-error on my side.
  • I have yet to find a place where my program needs to use LPP. I am not part of the operating system, and have no way to transport (and complement later) any specific information via these 8 bytes (as opposed to z/VM or z/OS which certainly do utilize the 3 fields (basic, logical, virtual) and would certainly produce confusing results with a program injecting its own 8 bytes).

The END


I coded it, commented out the QSI and LSCTL instructions, created entries as I expect the HW to do it, tested it, and removed the testcode…and then I learned that I have not read the fine print of the PTF (thanks to "E.", saving me hours of frustration in front of a z196) . Here is the fine print: 
Support for the CPU-Measurement Sampling Facility and ... interfaces for guest use are not provided.

So I can put this to a rest and wait. Maybe one day support for guests will show up.

Next

Next time I get a chance, I will write about CICS/TS and coding a simple TRUE and a GLUE.

1 comment:

  1. Meanwhile I had a chance to try it on a VSE running native on a z10 or higher. Here are the results, that do matter (a divide-exception because I forgot to handle the zero case is not worth mentioning nor is the priv-op-excep because I forgot to privilege me before doing LSCTL).
    1.) It works as documented (no trap - no fine print - nothing)
    2.) large memory objects in zVSE 5 (althru documented as if) does not support/utilise large pages (gives RC of 4) --- The code I prepared for testing under versions prior to 5 (4.1 thru 4.3) was used as elegant solution.

    ReplyDelete

Feel free to leave a comment or ask questions.