Non-Page Allocated Memory Sharing via Unified Architecture
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computing systems face inefficiencies in memory sharing between the CPU and GPU due to page alignment restrictions, leading to data corruption and reduced performance when handling non-page aligned memory allocations, as they typically only support page-granular memory sharing and lack efficient mechanisms for managing misaligned allocations.
Innovation Solution
A unified memory architecture (UMA) is implemented, allowing direct memory sharing between the CPU and GPU without data copying, using a device driver to manage and synchronize non-page aligned allocations, track allocation affinity, and prevent data corruption by delaying operations and managing misaligned page fragments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If page-granular memory sharing is used between CPU and GPU, then memory sharing is simplified and supported by operating system, but non-page aligned memory allocations cannot be shared efficiently leading to data corruption and reduced performance
Solution Approach 1:
The patent segments the memory management task by separating page-aligned memory sharing (handled by OS) from non-page aligned memory sharing (handled by driver). The driver manages misaligned allocations individually, dividing the memory sharing responsibility into two distinct layers to resolve the contradiction between simplicity and data integrity.
Solution Approach 2:
The graphics driver acts as an intermediary between the OS and GPU memory access. It translates GPU's non-page aligned memory requests into page-aligned operations that the OS can manage, while maintaining direct GPU access to misaligned allocations. This intermediary layer enables both simplicity and reliability.
2Productivity
If non-page aligned memory allocations are shared directly between CPU and GPU, then performance is improved by avoiding data copying, but data corruption occurs due to lack of synchronization mechanisms
Solution Approach 1:
The driver implements feedback mechanisms by tracking GPU access to memory regions and using this information to synchronize CPU and GPU operations. When the GPU accesses a memory region, the driver receives feedback and adjusts CPU access accordingly, preventing data corruption while maintaining high performance.
Solution Approach 2:
The driver performs preliminary synchronization actions before allowing CPU or GPU access to shared memory. It establishes proper access rights and synchronization barriers in advance, ensuring data consistency is maintained without requiring complex runtime synchronization during memory operations.
3Device complexity
If operating system manages all memory sharing at page granularity, then system simplicity is maintained, but fine-grained non-page aligned allocations cannot be shared even when GPU needs them
Solution Approach 1:
The patent segments memory management responsibilities: the OS handles page-aligned memory sharing (maintaining simplicity), while the graphics driver handles non-page aligned allocations (providing flexibility). This segmentation allows the system to maintain low complexity at the OS level while achieving high adaptability through driver-level management.
Solution Approach 2:
The graphics driver serves as an intermediary that translates GPU's fine-grained memory requests into operations the OS can manage. It bridges the gap between the OS's page-granular management and the GPU's need for fine-grained access, maintaining both simplicity and flexibility.
4Reliability
If data copying is used to enable GPU access to non-page aligned memory, then data integrity is maintained, but performance is reduced due to additional data transfer overhead
Solution Approach 1:
The patent uses a form of copying at the memory mapping level rather than data level. The driver creates separate memory mappings for CPU and GPU that both reference the same physical memory, enabling direct access without copying actual data. This maintains data integrity while avoiding the performance penalty of data transfer.
Data Source
AI summary
A method for sharing memory between a central processing unit (CPU) and an input/output (I/O) device of a computing device is described. The method may include creating an allocation of memory for the I/O device to operate on. The method includes detecting whether the allocation is not page-aligned, wherein an allocation is page-aligned when its base address and size be evenly divisible by the applicable page-size. The allocation may be successfully shared, even if not page-aligned, even if an operating system of the computing device doesn't support sharing of non-page-aligned allocations.