Dynamic techniques for optimizing soft real-time task performance in virtual machines
Methods are disclosed that dynamically improve soft real-time task performance in virtualized computing environments under the management of an enhanced hypervisor comprising a credit scheduler. The enhanced hypervisor analyzes the on-going performance of the domains of interest and of the virtualized data-processing system. Based on the performance metrics disclosed herein, some of the governing parameters of the credit scheduler are adjusted. Adjustments are typically performed cyclically, wherein the performance metrics of an execution cycle are analyzed and, if need be, adjustments are applied in a later execution cycle. In alternative embodiments, some of the analysis and tuning functions are in a separate application that resides outside the hypervisor. The performance metrics disclosed herein include: a “total-time” metric; a “timeslice” metric; a number of “latency” metrics; and a “count” metric. In contrast to prior art, the present invention enables on-going monitoring of a virtualized data-processing system accompanied by dynamic adjustments based on objective metrics.
1. A method comprising:
receiving, by a hypervisor in a data-processing system, a first metric that is associated with a first domain that is time-sensitive, wherein:
(i) the data-processing system comprises a first processor, a second processor, and a task queue comprising: a boost-priority region; an under-priority region; an over-priority region; and an idle-priority region, in decreasing priority value;
(ii) the first domain comprises a first plurality of tasks that were executed on the data-processing system;
(iii) the first metric is based on an amount of time, during an earlier time interval on the data-processing system, that was spent executing tasks of the first domain from the under-priority region of the task queue; and
(iv) a time-sensitive domain comprises tasks having an execution deadline;
receiving, by the data-processing system, a second metric that is associated with a second domain that is time-sensitive, wherein:
(i) the second domain comprises a second plurality of tasks that were executed on the data-processing system, and
(ii) the second metric is based on the amount of time, during the earlier time interval on the data-processing system, that was spent executing tasks of the second domain from the under-priority region of the task queue;
when the value of the first metric is below a first threshold and the value of the second metric is below a second threshold, bin-packing by the hypervisor the first plurality of tasks and the second plurality of tasks into at least a first bin-pack; and
executing by the first processor the tasks in the first bin-pack,
wherein:
the bin-packing is based on minimizing the number of processors that are to execute the first plurality of tasks and the second plurality of tasks;
the bin-packing is further based on providing a predetermined sufficient minimum processing time to tasks within the first bin-pack;
the first threshold is based upon a quality of performance of one or more tasks of the first plurality of tasks; and
the second threshold is based upon a quality of performance of one or more tasks of the second plurality of tasks.
2. The method of claim 1 wherein the hypervisor comprises a credit-scheduler.
3. The method of claim 1 wherein the application is separate from the hypervisor.
4. The method of claim 1 further comprising:
when an event arrives that is associated with a task that is time-sensitive, assign a boost priority to the task, wherein the assigning is independent of the pre-existing priority of the task, and
wherein the task is to be executed on the data-processing system in the later time interval in accordance with the boost priority.