none
How to set AllocationShrinkLoadRatio and AllocationGrowLoadRatio in HPC 2016 update 1 RRS feed

  • Question

  • Dear community,

    In the HPC 2008 release it seemed possible to adjust these trelholds in the broker config. In HPC 2016 these properties are part of the SessionStartInfoContract. However I'm not able to set the properties on code level. How can I set these values otherwise?

    Cheers,

    Joris


    Joris Cramwinckel


    Thursday, December 12, 2019 12:46 PM

Answers

All replies

  • Hi Joris,

    From the comment in SessionStartInfo, this configuration is obsolete and no longer take effect.

    Best regards,

    Zihao

    Thursday, January 2, 2020 1:58 AM
  • Thanks for the response Zihao, so as I understand there is no other way to influence the broker behavior? I wanted the client to be in control of the TargetResourceCount however if one sets the AdjustmentAllocationInterval to -1 the broker grow/shrink behavior is disabled hence the broker ignores TargetResourceCount adjustments from the client (see this thread). And setting the interval leads to both the broker and the client sending 'SetTargetResourceCount' events. Since the broker node decisions are based on the the Queue length, while the client is generating requests during the job these TargetResourceCount adjustments of the broker and the client can have opposing values, leading undesired shrinkage. 

    The root of my quest is that the initialization of a ServiceTask can take up to 40s and the request queue will sometimes be small or empty even-though more tasks will be generated during the job. This is why we want the client to manage to keep the service tasks warm. Any suggestions to achieve this properly are more then welcome. Unfortunately the option to mute the brokers involvement with the TargetResourceCount can not be enforced with the Load ratios.


    Joris Cramwinckel

    Thursday, January 2, 2020 12:00 PM
  • Hi Joris,

    Is setting longer AdjustmentAllocationInterval possible? If the the interval is set to a value longer than the waiting time expectation, when most "warmed up" service hosts should be able to survive automatic-shrink.

    Friday, January 3, 2020 2:16 AM
  • Hi Zihao,

    This is indeed the current solution. We put ourselves in the trade-off of keeping warmed-up service tasks alive versus not letting idle service tasks occupy a CPU for too long near the end of the job. Still, ideally we want to manage the brokers re-balancing manually, but there seems no robust way for doing this in our specific case.


    Joris Cramwinckel

    Monday, January 6, 2020 9:44 AM