Method for scheduling data flow task and apparatus

Patent metadata

Field	Value
Publication number	US-10558498-B2
Application number	US-201715598696-A
Country	US
Kind code	B2
Filing date	May 18, 2017
Priority date	Nov 19, 2014
Publication date	Feb 11, 2020
Grant date	Feb 11, 2020

Abstract

A method for scheduling a data flow task and an apparatus. The method includes: preprocessing a data flow task to obtain at least one subtask; classifying the subtask into a central processing unit (CPU) task group, a graphics processing unit (GPU) task group, or a to-be-determined task group; allocating the subtask to a working node; when the subtask belongs to the CPU task group, determining that a CPU executes the subtask; when the subtask belongs to the GPU task group, determining that a GPU executes the subtask; or when the subtask belongs to the to-be-determined task group, determining, according to costs of executing the subtask by a CPU and a GPU, a running platform (e.g., the CPU or the GPU) executes the subtask, where the cost includes duration of executing the subtask.

First claim

What is claimed is:

1. A method for scheduling a data flow task in a distributed heterogeneous system comprising at least one working node, the method comprising:
preprocessing the data flow task to obtain at least one subtask;
classifying a first subtask in the at least one subtask into a task group, wherein task groups of the distributed heterogeneous system comprise a central processing unit (CPU) task group, a graphics processing unit (GPU) task group, and a to-be-determined task group, wherein classifying a first subtask in the at least one subtask into a task group comprises:
classifying the first subtask into the CPU task group when the first subtask comprises indication information indicating that the first subtask is executed by a CPU; and
classifying the first subtask into the GPU task group when the first subtask comprises indication information indicating that the first subtask is executed by a GPU;
allocating the first subtask to a first working node in the at least one working node according to the task group to which the first subtask belongs and a resource status of the at least one working node;
determining that a CPU corresponding to the first working node executes the first subtask when the first subtask belongs to the CPU task group;
determining that a GPU corresponding to the first working node executes the first subtask when the first subtask belongs to the GPU task group; and
determining, according to costs of executing the first subtask by a CPU and a GPU, a running platform that executes the first subtask when the first subtask belongs to the to-be-determined task group, wherein each of the costs comprises duration of executing the first subtask.

2. The method according to claim 1 , wherein classifying a first subtask in the at least one subtask into a task group comprises:
estimating first duration of executing the first subtask by the CPU and second duration of executing the first subtask by the GPU when the first subtask does not comprise indication information, and classifying the first subtask into a task group according to the first duration and the second duration.

3. The method according to claim 2 , wherein classifying the first subtask into a task group according to the first duration and the second duration comprises:
classifying the first subtask into the CPU task group if a ratio of the first duration to the second duration is less than a first preset threshold;
classifying the first subtask into the GPU task group if the ratio of the first duration to the second duration is greater than a second preset threshold; and
classifying the first subtask into the to-be-determined task group if the ratio of the first duration to the second duration is not less than the first preset threshold and is not greater than the second preset threshold.

4. The method according to claim 1 , further comprising:
recording execution log information of the first subtask into a performance database, wherein the execution log information comprises a data volume of the first subtask, required waiting duration before the first subtask is executed, and a running platform and running duration of the first subtask.

5. The method according to claim 4 , further comprising:
querying the performance database, and calculating first average duration of executing subtasks by the CPU corresponding to the first working node and second average duration of executing the subtasks by the GPU corresponding to the first working node; and
adjusting, according to the first average duration and the second average duration, subtasks distribution on the first working node.

6. A method for scheduling a data flow task in a distributed heterogeneous system comprising at least one working node, the method comprising:
preprocessing the data flow task to obtain at least one subtask;
classifying a first subtask in the at least one subtask into a task group, wherein task groups of the distributed heterogeneous system comprise a central processing unit (CPU) task group, a graphics processing unit (GPU) task group, and a to-be-determined task group;
allocating the first subtask to a first working node in the at least one working node according to the task group to which the first subtask belongs and a resource status of the at least one working node;
determining that a CPU corresponding to the first working node executes the first subtask when the first subtask belongs to the CPU task group;
determining that a GPU corresponding to the first working node executes the first subtask when the first subtask belongs to the GPU task group;
determining, according to costs of executing the first subtask by a CPU and a GPU, a running platform that executes the first subtask when the first subtask belongs to the to-be-determined task group, wherein each of the costs comprises duration of executing the subtask;
calculating first average duration of executing subtasks by the CPU corresponding to the first working node and second average duration of executing the subtasks by the GPU corresponding to the first working node; and
adjusting, according to the first average duration and the second average duration, subtasks distribution on the first working node;
wherein adjusting, according to the first average duration and the second average duration, subtasks distribution on the first working node comprises:
allocating first O subtasks that are in the to-be-determined task group on the first working node and that have greatest acceleration ratios to the GPU task group on the first working node if the first average duration is greater than the second average duration, wherein the acceleration ratio is a ratio of a time of executing a subtask by the CPU to a time of executing the subtask by the GPU; and
allocating first P subtasks that are in the to-be-determined task group on the first working node and that have smallest acceleration ratios to the CPU task group on the first working node if the first average duration is less than the second average duration, wherein O and P are positive integers.

7. The method according to claim 5 , wherein adjusting, according to the first average duration and the second average duration, subtasks distribution on the first working node comprises:
allocating first O subtasks that are in the CPU task group on the first working node and that have greatest acceleration ratios to the to-be-determined task group on the first working node if the first average duration is greater than the second average duration, wherein the acceleration ratio is a ratio of a time of executing a subtask by the CPU to a time of executing the subtask by the GPU; and
allocating first P subtasks that are in the GPU task group on the first working node and that have smallest acceleration ratios to the to-be-determined task group on the first working node if the first average duration is less than the second average duration, wherein O and P are positive integers.

8. An apparatus for use in a distributed heterogeneous system comprising at least one working node, the apparatus comprising:
a memory configured to store instructions; and
a processor coupled to the memory and configured to execute the instructions to:
preprocess a data flow task to obtain at least one subtask;
classify a first subtask in the at least one subtask into a task group, wherein task groups of the distributed heterogeneous system comprise a central processing unit (CPU) task group, a graphics processing unit (GPU) task group, and a to-be-determined task group, wherein to classify a first subtask in the at least one subtask into a task group comprises:
classify the first subtask into the CPU task group when the first subtask comprises indication information indicating that the first subta

Assignees

• Huawei Tech Co Ltd

Inventors

• Deng Liqun
• Zhu Junhua

Classifications

• G06F9/5027 Primary

  the resource being a machine, e.g. CPUs, Servers, Terminals · CPC title

• G06F9/4881

  Scheduling strategies for dispatcher, e.g. round robin, multi-level priority queues · CPC title

• G06F9/5044 Primary

  considering hardware capabilities · CPC title