CP Overhead

This system has a 6 linux guests and very high cpu overhead. In the following one minute sample (that is repeated every minute every day), there is one line of data per physical processor. There are several indications of the problem, that without proper tools, would be difficult to understand. Seeing data from other systems with CPU performance problems may be enlightening.

This shows in order, CPU Total utilization on the processor User overhead - system time assigned to specific users System overhead - CPU time than can not be assigned. This number is normally in the 2-5 percent. Diagnose Instructions per second (Yup, this is VERY large), Instruction Simulation per second (Same as diagnose), SIE Interceptions per second - this really very bad. Each SIE interception requires CP involvement. SIE fast path - where CP could take a short code path

  •  

  • This is a sure sign of a problem program that needs to be fixed.

    ESACPUA
    --------------------------------------------
        <--Internal (per seco
         Totl  Ovrhead  Diag  Inst     SIE  Fast
    CPU  Util  Usr Sys  nose   Sim  intrcp  path
    ---  ----  --- ---  ----  ----  ------  ----
    0    97.3   31  22  68K   68K   69220   2.42
    1    97.4   30  21  69K   69K   69652   3.65
    2    97.4   30  21  69K   69K   69478   3.02
         ----  --- ---  ----  ----  ------  ----
          292   92  65  206K  206K  208350  9.08

    The following report fragment shows the dispatching activity on each of the 3 processors. If there are master processor constraints, they show up here - and in this case there are none. However, note the large number of dispatches per second in the last column? Pretty much the same as the SIE intercepts per second. This Proves that indeed, there is CP dispatching overhead every time the linux guest issues a diagnose.

    This shows, in order, VMDBK Steals: when a processor completes it's work, it goes looking for another VMDBK. Moves per second to Master: Again normal, will be high when there is VM Master precessor dependencies. PLDV Lengths: Number of users ready to run on each processor. Dispatch Long Paths: This number is excessive. It is the number of times a VMDBK was dispatched. Numbers in the 100 range is typical. This is a source of high overhead.

    ESAPLDV
            <--------PLDV Lengths-------> Dispatcher
    CPU  Steals  To Master    Avg  Max  Mstr MstrMax %Empty Long Paths
     -   ------  ---------   ----  ---  ---- ------- ------ ----------

     0    823.8        0.7    0.2  1.0     .       .   83.3    70489.8
     1    111.1          0    0.4  1.0     .       .   55.0    70454.2
     2    196.4          0    0.4  2.0     .       .   61.7    70056.8
         ------  ---------   ----  ---  ---- ------- ------ ----------
         1131.3        0.7    1.0  4.0     .       .  200.0   211000.7