Apparatus and method for efficient estimation of the energy dissipation of processor based systems
A system and method of scheduling tasks, comprising receiving activity and performance data from registers or storage locations maintained by hardware and an operating system; storing calibration coefficients associated with the activity and performance data; computing an energy dissipation rate based on at least the activity and performance data; and scheduling tasks under the operating system based on the computed energy dissipation rate.
1 . A method of workload scheduling, comprising:
determining an energy budget for each of a plurality of racks in which a plurality of servers reside;
maintaining quality of service requirements for each of a plurality of different types of tasks;
classifying each task and determining an applicable quality of service requirement for execution of the respective task;
predicting an impact of each task on an energy consumption of a server and rack upon which it executes; and
scheduling a task for execution on a respective server in a respective rack in dependence on the classification of the task, the predicted impact of the task on energy consumption of the server and rack on which it executes, the energy budget for the rack, and the quality of service requirements, to meet the quality of service for the task.
2 . The method according to claim 1 , wherein the task is scheduled for execution on a server in a rack having high utilization in preference to a server in a rack with low utilization.
3 . The method according to claim 1 , wherein each server has an active state in which is processes tasks, and an inactive state which does not process tasks.
4 . The method according to claim 3 , wherein the scheduler is configured to control the active and inactive state of the server, and to selectively schedule the task to a respective server on a rack near maximum energy budget, to maximize a number of racks having only inactive servers.
5 . The method according to claim 1 , wherein tasks having common classifications are executed on a common server.
6 . The method according to claim 1 , wherein each task is classified according to whether it is CPU bound, disk-bound, or network-bound.
7 . The method according to claim 6 , wherein the server has a variable performance controllable in dependence on a control signal, and the scheduler schedules tasks having respectively similar classification for execution on a server back-to-back without intervening changes in the control signal.
8 . The method according to claim 1 , further comprising scheduling a cooling of a rack in dependence on a predicted thermal condition of the rack dependent on tasks to be executed on servers on the rack.
9 . The method according to claim 1 , further comprising executing software on each server configured to access to registers and counters of server hardware, to track an activity of the server.
10 . The method according to claim 9 , wherein the software comprises a multitasking operating system of the server which executes the task.
11 . The method according to claim 1 , further comprising revising the energy budget for a respective server in dependence on a task processing workload allocated to the respective server.
12 . The method according to claim 1 , wherein the scheduling is further dependent on an amount of time predicted to change a performance state of a respective server.
13 . The method according to claim 1 , further comprising maintaining data tables that record energy/performance characteristics of each rack, utilization statistics for each rack, and an environmental temperature of each rack, wherein the scheduling of the task is further dependent on the data tables.
14 . A method of workload scheduling, comprising:
determining an energy budget for each of a plurality of virtual machines, an energy budget for each of a plurality of servers on which the virtual machines are organized, and an energy budget for each of a plurality of racks into which the plurality of servers reside, wherein each server has an active state for processing tasks, and an inactive state in which the server does not process tasks;
maintaining a variety of data tables that record energy/performance characteristics of each rack, rack utilization statistics, and an environmental temperature of the plurality of racks;
maintaining quality of service requirements for each of a plurality of different types of tasks;
classifying each newly received task;
predicting an impact of an incoming task processing workload on energy consumption of a server and rack upon which is executes; and
scheduling the newly received task for execution by a respective virtual machine in dependence on the characteristics of the newly received task, the predicted impact of the incoming task processing workload, the energy budget for the plurality of virtual machines, the energy budget for the plurality of servers, the energy budget for the plurality of racks, the data tables, and the quality of service requirements, to maximize a workload allocated to each active server before activating an additional server from the inactive state, while meeting the quality of service for the newly received task.
15 . The method according to claim 14 , further comprising migrating a respective virtual machine from a first server in a first rack to a second server in a second rack in dependence on the energy budget of the first rack and the energy budget of the second rack.
16 . The method according to claim 14 , wherein the plurality of racks are in a room, wherein the scheduling is further dependent on a predicted hotspot within the room dependent on the scheduling.
17 . The method according to claim 14 , wherein the scheduling is performed by a hierarchical control system comprising a first control component for controlling each server, a second control component for controlling each rack, and a third control component for controlling a plurality of spatially proximate racks.
18 . A system for workload scheduling, comprising:
a memory configured to store an energy budget for each of a plurality of racks in which a plurality of servers reside, utilization statistics for each rack, and an environmental temperature of each rack, and a quality of service requirement for each of a plurality of different types of tasks;
at least one processor configured to
classify each task, and determine an applicable quality of service requirement for execution of the respective task;
predict an impact of each task on an energy consumption of a server and rack upon which is executes; and
schedule a task for execution on a respective server in a respective rack in dependence on the classification of the task, the predicted impact of the task on energy consumption of the server and rack on which it executes, the energy budget for the rack, and the quality of service requirements, to meet the quality of service for the task; and
an output port configured to communicate the tack schedule.
19 . The system according to claim 18 , wherein:
each server has an active state in which is processes tasks, and an inactive state which does not process tasks; and
the scheduler is configured to control the active and inactive state of the server, and to selectively schedule the newly received task to a respective server on a rack near maximum energy budget, to maximize a number of racks having only inactive servers
wherein the task is scheduled for execution on a server in a rack having high utilization statistics in preference to a server in a rack with low utilization statistics.
20 . The system according to claim 18 , wherein:
each server has a variable performance controllable in dependence on a control signal;
each task is classified according to whether it is CPU bound, disk-bound, or network-bound;
tasks having common classifications are scheduled for execution on a common server; and
the scheduler is configured to schedule tasks having respectively similar classification for execution on a server back-to-back without intervening changes in the control signal.