Dynamic computational acceleration using a heterogeneous hardware infrastructure

A computing system, comprising:

• at least one processing unit;
• and
• a non-transitory computer-readable storage medium storing instructions that, when executed, cause the at least one processing unit to:
• in a distributed computing environment comprising a plurality of distributed computing systems that are communicatively coupled to one another via at least one network, identify, during execution of a software application, platform-independent instructions that are configured to perform at least one computational task associated with execution of the software application, wherein the plurality of computing systems includes a heterogeneous group of specialized processing units, wherein the platform-independent instructions have a platform-independent format that is not specific to any particular processing unit of the heterogeneous group of specialized processing units, wherein the software application is associated with an application executable previously compiled from application source code using a first compiler, wherein the software application is further associated with the platform-independent instructions that, prior to execution of the software application, were compiled from computational source code using a second compiler, and wherein the platform-independent instructions are not native to or executable by any particular processing unit of the heterogeneous group of specialized processing units;
• determine real-time performance information to collect and store during performance of the at least one computational task by the heterogeneous group of specialized processing units, wherein the real-time performance information indicates an amount of one or more constrained resources utilized by one or more of the specialized processing units while performing the at least one computational task during execution of platform-dependent instructions that are specific to each respective specialized processing unit;
• after identifying the platform-independent instructions, determine one or more scheduling criteria that are associated with the platform-independent instructions, wherein the one or more scheduling criteria are based at least in part on the real-time performance information previously collected while performing the at least one computational task by the one or more of the specialized processing units during prior execution of platform-dependent instructions specific to each respective specialized processing unit;
• select, from the heterogeneous group of specialized processing units and based on the one or more scheduling criteria, a specialized processing unit to perform the at least one computational task, wherein selecting the specialized processing unit comprises selecting one of the heterogeneous group of specialized processing units that utilizes a relatively fewer amount of the one or more constrained resources, with respect to one or more other specialized processing units in the group, when performing the at least one computational task during execution of platform-dependent instructions specific to the selected specialized processing unit;
• during execution of the software application, convert the platform-independent instructions into the platform-dependent instructions that are specific to the selected specialized processing unit, wherein the application executable includes references, provided by one or more libraries, to the platform-independent instructions that are resolved at runtime during execution of the software application, and wherein the platform-dependent instructions comprise optimized native code executable by the selected specialized processing unit;
• and
• provide the platform-dependent instructions to the selected specialized processing unit for execution in order to perform the at least one computational task.

A non-transitory computer-readable storage medium comprising instructions that, when executed, cause a computing system to perform operations comprising:

• in a distributed computing environment comprising a plurality of distributed computing systems that are communicatively coupled to one another via at least one network, identifying, during execution of a software application, platform-independent instructions that are configured to perform at least one computational task associated with execution of the software application, wherein the plurality of computing systems includes a heterogeneous group of specialized processing units, wherein the platform-independent instructions have a platform-independent format that is not specific to any particular processing unit of the heterogeneous group of specialized processing units, wherein the software application is associated with an application executable previously compiled from application source code using a first compiler, wherein the software application is further associated with the platform-independent instructions that, prior to execution of the software application, were compiled from computational source code using a second compiler, and wherein the platform-independent instructions are not native to or executable by any particular processing unit of the heterogeneous group of specialized processing units;
• determining real-time performance information to collect and store during performance of the at least one computational task by the heterogeneous group of specialized processing units, wherein the real-time performance information indicates an amount of one or more constrained resources utilized by one or more of the specialized processing units while performing the at least one computational task during execution of platform-dependent instructions that are specific to each respective specialized processing unit;
• after identifying the platform-independent instructions, determining one or more scheduling criteria that are associated with the platform-independent instructions, wherein the one or more scheduling criteria are based at least in part on the real-time performance information previously collected while performing the at least one computational task by the one or more of the specialized processing units during prior execution of platform-dependent instructions specific to each respective specialized processing unit;
• selecting, from the heterogeneous group of specialized processing units and based on the one or more scheduling criteria, a specialized processing unit to perform the at least one computational task, wherein selecting the specialized processing unit comprises selecting one of the heterogeneous group of specialized processing units that utilizes a relatively fewer amount of the one or more constrained resources, with respect to one or more other specialized processing units in the group, when performing the at least one computational task during execution of platform-dependent instructions specific to the selected specialized processing unit;
• during execution of the software application, converting the platform-independent instructions into the platform-dependent instructions that are specific to the selected specialized processing unit, wherein the application executable includes references, provided by one or more libraries, to the platform-independent instructions that are resolved at runtime during execution of the software application, and wherein the platform-dependent instructions comprise optimized native code executable by the selected specialized processing unit;
• and
• providing the platform-dependent instructions to the selected specialized processing unit for execution in order to perform the at least one computational task.