System and method for increasing robustness of heterogeneous computing systems
Disclosed is a method for task pruning that can be utilized in existing resource allocation systems to improve the systems' robustness without requiring changing to existing mapping heuristics. The pruning mechanism leverages a probability model, which calculates the probability of a task competing before its deadline in the presence of task dropping, and only schedules tasks that are likely to succeed. Pruning tasks whose chance of success is low improves the chance of success for other tasks. Tasks that are unlikely to succeed are either deferred from current scheduling event or are preemptively dropped from the system. The pruning method can benefit service providers by allowing them to utilize their resources more efficiently and use them only for tasks that can meet their deadlines. The pruning method further helps end users by making the system more robust in allowing more tasks to complete on time.
1. A method for performing task pruning in a resource allocation system,
wherein the resource allocation system comprises one or more mapping heuristics;
wherein one or more tasks arrive into the resource allocation system;
wherein the one or more tasks are queued upon arrival into a batch queue;
comprising:
establishing a dropping threshold;
dropping one or more low quality tasks by a pruning mechanism;
creating a mapping event;
evaluating the oversubscription level of the computing system;
for each task in the batch queue, the mapping event attempts to map said tasks in the batch queue to one or more machine queues, comprising:
creating a virtual queue of one or more machine-task mappings;
calculating a completion time distribution of one or more unmapped tasks to one or more machines;
deferring tasks with low chances of success; and
returning the tasks that have been deferred to the batch queue until a subsequent mapping event;
calculating an impact of dropping a low quality task, wherein a pruning mechanism calculates a completion time and a probability of timely completing the task subsequent to the dropped low quality task;
determining a best machine for each task by the mapping heuristic;
selecting which task will be paired with which machine queue by the mapping heuristic; and
mapping each task from the batch queue to a machine queue until no tasks remain in the batch queue;
wherein calculating the impact further comprises generating one or more impulses representing a completion time of each task.
2. The method of claim 1 , wherein low quality tasks comprise tasks that have missed their deadlines.
3. The method of claim 1 , wherein low quality tasks comprise tasks that have a low chance of success, wherein tasks with a low chance of success are those that are unlikely to be completed before their deadlines.
4. The method of claim 1 , wherein a mapping event occurs when a task completes its execution.
5. The method of claim 1 , wherein a mapping event occurs when a new task arrives into the system.
6. The method of claim 1 , wherein the tasks assigned to a machine queue are processed in a first come first served manner.
7. The method of claim 1 , where the tasks assigned to a machine queue are executed in isolation without preemption or multitasking.
8. The method of claim 1 , wherein the dropping threshold is dynamically adjusted based upon a skewness and a position in the batch queue of each task.
9. The method of claim 1 , wherein the pruning mechanism's level of aggression is dynamically adjusted in reaction to the oversubscription level in the computing system.
10. The method of claim 1 , wherein the oversubscription level is determined by evaluating a weight average number of the tasks that missed their deadlines during past mapping events.
11. The method of claim 1 , wherein at least one mapping heuristic is configured to leverage the pruning mechanism to maximize robustness.
12. The method of claim 1 , wherein at least one mapping heuristic is configured to achieve fairness across the tasks.
13. The method of claim 1 , wherein best machine for the task is the machine with the highest robustness.
14. A pruning mechanism for a resource allocation system comprising:
an accounting module;
a toggle module;
a pruner module;
a pruning configuration; and
a fairness module;
wherein the accounting module is configured to gather one or more task's information from the resource allocation system;
wherein the toggle module comprises functionality to measure the heterogeneous computing system's oversubscription level;
wherein the fairness module comprises functionality to avoid bias; and
wherein the toggle module measures the computing system's oversubscription level by:
monitoring a total amount of tasks that have missed their deadlines since a prior mapping event; and
identifying the computing system as oversubscribed if the total amount of tasks exceeds a Dropping Toggle.
15. The pruning mechanism of claim 14 , wherein the pruning configuration comprises a user-defined threshold required level of success for tasks.
16. The pruning mechanism of claim 14 , wherein the fairness module avoids bias by:
monitoring a one or more task types that are dropped by the pruning mechanism; and
adjusting the pruning mechanism to avoid bias against task types that are consistently dropped.
17. The pruning mechanism of claim 14 , wherein the accounting module further comprises capability to record the tasks that are completed or dropped.