GPU-shared dispatching and single-machine multi-card methods, systems, and devices

A Graphics Processing Unit (GPU)-shared single-machine multi-card method, being applied to a target controlled host and comprising:

• in response to a determination that a current pod runs, setting a resource occupation marker in annotation of the current pod to identify GPU information used during the current pod running through the resource occupation marker;
• receiving a resource division request sent by at least one of a plurality of pods bound to the target controlled host;
• traversing the plurality of pods bound to the target controlled host to determine whether there are a plurality of preset pods recording first resource division information consistent with second resource division information recorded in the resource division request;
• in response to a determination that there is only one of the plurality of preset pods recording the first resource division information consistent with the second resource division information recorded in the resource division request, determining the one of the plurality of preset pods as a target pod;
• in response to a determination that there are the plurality of preset pods recording the first resource division information consistent with the second resource division information recorded in the resource division request, screening out the target pod without the resource occupation marker from the plurality of preset pods;
• writing GPU information of a target GPU, written to an annotation of the target pod by a dispatcher of a central control host, to an environment variable of the target pod;
• registering one or more corresponding virtual display cards by use of resource division information in the target pod, and correspondingly dividing a frame buffer of the target GPU to each of the one or more corresponding virtual display card;
• and
• setting the resource occupation marker for the target pod to indicate that the target GPU recorded in the annotation of the target pod is already occupied.

A Graphics Processing Unit (GPU)-shared dispatching device, comprising:

• a memory, configured to store a computer program;
• and
• a processor, configured to execute the computer program to:
• in response to a determination that a current pod runs, set a resource occupation marker in annotation of the current pod to identify first GPU information used during the current pod running through the resource occupation marker;
• in response to a determination that the first GPU information of the current pod is updated, set an update marker for the current pod;
• query first GPU information in an environment variable of a non-updated pod in each of a plurality of controlled hosts by use of the resource occupation marker and the update marker of a pod, the non-updated pod being a pod that is already running but yet does not update the first GPU information;
• update the first GPU information to an annotation of the non-updated pod, and add the update marker for the non-updated pod;
• screen out a dispatchable controlled host without an unmarked pod from the plurality of controlled hosts, the unmarked pod being a pod without the update marker;
• screen out, by use of state information of a GPU in the dispatchable controlled host, a target controlled host satisfying a first preset condition from the dispatchable controlled host and a target GPU satisfying a second preset condition in the target controlled host;
• write second GPU information of the target GPU to an annotation of a pod to be allocated;
• and
• allocate the pod to be allocated to the target controlled host.

A Graphics Processing Unit (GPU)-shared single-machine multi-card device, comprising:

• a memory, configured to store a computer program;
• and
• a processor, configured to execute the computer program to:
• in response to a determination that a current pod runs, set a resource occupation marker in annotation of the current pod to identify GPU information used during the current pod running through the resource occupation marker;
• receive a resource division request sent by at least one of a plurality of pods bound to the target controlled host;
• traverse the plurality of pods bound to the target controlled host to determine whether there are a plurality of preset pods recording first resource division information consistent with second resource division information recorded in the resource division request;
• in response to a determination that there is only one of the plurality of preset pods recording the first resource division information consistent with the second resource division information recorded in the resource division request, determine the one of the plurality of preset pods as a target pod;
• in response to a determination that there are the plurality of preset pods recording the first resource division information consistent with the second resource division information recorded in the resource division request, screen out the target pod without the resource occupation marker from the plurality of preset pods;
• write GPU information of a target GPU, written to an annotation of the target pod by a dispatcher of a central control host, to an environment variable of the target pod;
• register one or more corresponding virtual display cards by use of resource division information in the target pod, and correspondingly divide a frame buffer of the target GPU to each of the one or more corresponding virtual display card;
• and
• set the resource occupation marker for the target pod to indicate that the target GPU recorded in the annotation of the target pod is already occupied.