On-Demand Kernel Assessment for Heterogeneous Cloud Accelerators
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Cloud management systems face challenges in selecting computing nodes with appropriate accelerators for applications due to the heterogeneity of hardware and software environments, and existing profiling tools add significant overhead and are not suitable for production environments.
Innovation Solution
A framework for generating kernel assessment applications that can be deployed to collect insights on hardware-accelerated applications, allowing for on-demand, automated, and scalable evaluation of kernels across different server nodes, independent of the main application.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional profiling tools are used to evaluate application kernels, then detailed performance insights can be obtained, but significant overhead is introduced and production environments are disrupted
Solution Approach 1:
The patent segments the kernel evaluation process into a separate assessment framework that operates independently from the main application execution. The assessment framework creates dedicated assessment applications that can be executed separately to gather kernel insights without interfering with the production application's normal operation, thus obtaining measurement precision without significant overhead.
Solution Approach 2:
The patent introduces an intermediary assessment framework that acts as a mediator between the production application and the evaluation process. This framework creates assessment applications that serve as intermediaries to collect kernel insights, allowing performance measurement without directly profiling the main application during its execution, thereby avoiding disruption to production environments.
2Adaptability or versatility
If hardware accelerators from different vendors are used, then computing flexibility and versatility are improved, but hardware heterogeneity makes kernel selection and resource allocation more difficult
Solution Approach 1:
The patent changes the parameters of kernel assessment by creating assessment applications that can be executed on different hardware accelerators with varying configurations. By systematically varying execution parameters across different accelerator types and collecting performance data, the framework builds comprehensive profiles that simplify the resource allocation decision-making process despite hardware heterogeneity.
Solution Approach 2:
The patent creates copies of the original application in the form of assessment applications that can be executed on different hardware accelerators. These assessment applications serve as replicas that allow evaluation of kernel performance across various accelerator types without affecting the original application, enabling comprehensive comparison and simplified selection based on collected metrics.
3Manufacturing precision
If comprehensive kernel assessment is performed on all possible hardware configurations, then accurate resource allocation can be achieved, but the assessment process becomes time-consuming and scalable only with difficulty
Solution Approach 1:
The patent performs preliminary actions by pre-creating and caching assessment application templates that can be quickly instantiated for different hardware configurations. This preliminary preparation allows the system to rapidly assess kernels on various accelerators without time-consuming setup procedures each time, achieving both accurate resource allocation and scalability through efficient reuse of assessment frameworks.
Data Source
AI summary
A method and system of assessment of applications in a cloud infrastructure includes generating a kernel assessment application to analyze an application to be executed on the cloud infrastructure, deploying the kernel assessment application in the cloud infrastructure, executing the kernel assessment application in the cloud infrastructure, and storing kernel insights collected from the kernel assessment application to be utilized for executing the application in the cloud infrastructure.