GPU Operator Pushdown in Big Data Caches to Cut Data Migration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The separation of storage and computing resources in big data clusters leads to increased data migration requirements and deteriorated data processing performance, necessitating a solution to reduce data migration costs and improve efficiency.
Innovation Solution
A data processing system utilizing a multi-level operator pushdown approach with GPUs in a cache acceleration and storage layer, optimizing scheduling and execution of operators based on real-time resource usage to minimize data movement and leverage heterogeneous computing power.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If storage and computing are separated into independent clusters, then resource utilization flexibility is improved, but data migration requirements increase and processing performance deteriorates
Solution Approach 1:
The system segments the data processing architecture into three independent subsystems: engine layer (first subsystem), cache acceleration layer (second subsystem), and storage layer (third subsystem). Each subsystem can be independently configured and optimized, allowing flexible resource allocation while maintaining efficient data flow through defined interfaces and protocols.
Solution Approach 2:
The cache acceleration layer acts as an intermediary between the engine layer and storage layer. It buffers data temporarily, reducing direct data migration between computing and storage clusters. The cache subsystem manages data retrieval and pre-processing, minimizing the data transfer burden on both storage and computing resources.
2Speed
If GPUs are introduced for operator execution, then processing speed is improved, but system complexity increases
Solution Approach 1:
The cache acceleration layer serves as an intermediary that manages GPU resources. It converts operators into intermediate representation operators suitable for GPU execution, handles scheduling, and manages the complexity of GPU operations. This shields the engine layer from direct GPU complexity while enabling accelerated processing.
Solution Approach 2:
The system changes the execution parameter from CPU-based operator execution to GPU-based execution. The cache acceleration layer transforms operators into GPU-compatible intermediate representations, enabling parallel processing and significantly improving execution speed for computationally intensive operations.
3Productivity
If data is pushed down to storage layer for GPU execution, then processing efficiency is improved, but data migration costs increase
Solution Approach 1:
Instead of migrating all data to the storage layer for GPU execution, the system selectively pushes down only those operators that benefit from GPU acceleration. The cache acceleration layer determines which operators should be executed on GPUs based on their computational characteristics, avoiding unnecessary data migration for operators that don't require GPU processing.
Solution Approach 2:
The cache acceleration layer performs preliminary processing and filtering of data before it reaches the storage layer. By pre-processing data in the cache, the system reduces the volume of data that needs to be migrated to storage for GPU execution, thereby improving efficiency while minimizing migration costs.
Data Source
AI summary
A data processing system includes a first subsystem implementing an engine layer, a second subsystem implementing a cache acceleration layer, and a third subsystem implementing a storage layer. The cache acceleration layer and the storage layer include GPUs. The first subsystem is configured to determine primitive operators to be executed by the GPUs and a scheduling plan of the primitive operators based on a query request, and output the scheduling plan to the second subsystem. The second subsystem converts the primitive operators into intermediate representation operators and schedules the intermediate representation operators to second execution objects based on the scheduling plan. The second subsystem drives, using a concurrency model, third execution objects to execute the intermediate representation operators. Execution results are output by the third execution objects to the first subsystem, and the execution results are used to obtain a query result.