Many-Core Accelerator Memory Management for Throughput
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The shortage of memory areas in many-core accelerators can lead to errors and decreased server throughput due to longer memory reserving and freeing processing times compared to host processors, especially when multiple applications are executed simultaneously.
Innovation Solution
An information processing device with processors that manage memory requests by determining whether requested memory areas can be reserved or freed, leaving pending requests when necessary to avoid errors and optimize memory usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If memory reserving and freeing processing is performed by OS in response to requests from applications, then memory management is achieved, but processing time increases and server throughput decreases
Solution Approach 1:
The patent extracts the memory reserving and freeing processing functions from the OS and implements them directly in the many-core accelerator. This allows the accelerator to manage its own memory resources independently, eliminating the time-consuming interaction with the OS and thereby improving server throughput while maintaining correct memory management.
Solution Approach 2:
The many-core accelerator is designed to autonomously manage its own memory allocation and deallocation without requiring OS intervention. The accelerator includes internal logic to track memory usage, allocate memory from its own pool, and free memory when no longer needed, enabling self-service memory management that reduces processing delays.
2Reliability
If memory area is reserved for many-core accelerator, then operation processing can be performed, but memory shortage may occur causing application errors
Solution Approach 1:
The patent implements dynamic memory management in the many-core accelerator that allows memory to be allocated and freed based on actual application needs. The accelerator tracks memory usage and can release memory when applications no longer need it, making the available memory pool dynamic rather than static, thus preventing memory shortage errors while maintaining necessary memory for ongoing operations.
3Productivity
If multiple applications are executed simultaneously, then server processing capacity is improved, but memory resource contention increases leading to longer processing times
Solution Approach 1:
The patent segments the memory management functions by giving each many-core accelerator its own independent memory pool and management logic. This segmentation allows multiple accelerators to manage memory independently without contending for the same memory resources or requiring coordinated allocation, thereby reducing processing time even when multiple applications run simultaneously across multiple accelerators.
Data Source
AI summary
The present invention discloses an information-processing device with which it is possible to improve throughput by avoiding application stoppage due to the shortage of a memory area, and by reducing the application processing time needed for reserving and freeing memory. This information-processing device is provided with: a memory-processing control unit configured to, when a memory-reserving request for a memory is received, determine whether an area size requested by the memory-reserving request can be reserved in a free area of the memory, and when a memory-freeing request for the memory is received, determines whether there is the memory-reserving request that is left pending; a reserving request processing unit configured to leave the memory-reserving request pending when it is determined by the memory-processing control unit that the area size cannot be reserved; and a freeing request processing unit which destroys or leaves pending the memory-freeing request when it is determined by the memory-processing control unit that there is no memory-reserving request that is left pending.