Forum Discussion

DarioGB_339840's avatar
DarioGB_339840
Icon for Altostratus rankAltostratus
May 14, 2018

HMS CPU Usage

Hello folks!

 

I wondering if someone could help me with this:

 

How Host Management Subsystem uses CPU resources in the new version of TMOS (v11.5+)?

 

I have reviewed this topic (https://devcentral.f5.com/questions/tmm-and-hms-hardware-resources) but I think this applies to a very old TMOS and the behaviour could have been changed in new versions (11.5+).

 

Base on this https://support.f5.com/csp/article/K15003, new version of TMOS uses CPU resources as next:

 

  • odd-numbered hyper-threads are used to data-plane tasks
  • even-numbered hyper-trheads are used to control-plane tasks

1) First question -> Are Management tasks equivalent to control-tasks? (i.e. TMSH, SNMP, REST are processed in even-numbered threads?)

 

2) Second question -> In case of just a single CPU is used (eq to old behaviour), Does HMS share CPU resources with TMM? Which CPU number is used?

 

Thanks for your help.

 

KR, Dario.

 

application delivery
TMOS

1 Reply

Sort By
  • Simon_Blakely's avatar
    Simon_Blakely
    Icon for Employee rankEmployee
    May 15, 2018

    odd-numbered hyper-threads are used to data-plane tasks

     

    even-numbered hyper-trheads are used to control-plane tasks

     

    Other way round - even vCPUs are data-plane, odd vCPUs are control-plane

     

    1) First question -> Are Management tasks equivalent to control-tasks? (i.e. TMSH, SNMP, REST are processed in even-numbered threads?)

     

    Yes. Those processes run on the odd-numbered cores.

     

    2) Second question -> In case of just a single CPU is used (eq to old behaviour), Does HMS share CPU resources with TMM? Which CPU number is used?

     

    A single non-hyperthreaded core will behave in the old way, with control-plane and data-plane processes sharing CPU resource.

     

    A single hyperthreaded CPU core will be split, and tmm (data-plane) runs on core0 and control-plane processes run on core1.

     

    There is an option in VCMP to assign a single hyperthread core to a guest, but this does have some limitations.

     

    Our development team estimates that for typical LTM workloads, a single CPU core using HT-Split can safely reach about 120% of the performance of the single (non-hyperthreaded) CPU core.

     

    Once core0 reaches 80% of the single non-hyperthreaded CPU core performance, core1 will be limited to 20%. Core0 performance can then approach 100% (with the two hyperthreads reaching about 120% utilization).