NUMA Memory Zeroing via Affinity-Aware Parallel Worker Threads
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for zeroing large memory regions in computing systems with non-uniform memory access (NUMA) are inefficient, leading to prolonged teardown times and visible delays, especially as memory sizes increase, due to the time-consuming nature of traditional zeroing processes and the impact of varying access pathways between processors and memory locations.
Innovation Solution
The implementation of a method that sorts memory pages by affinity domain and utilizes worker threads local to each domain for parallel zeroing, allowing memory to be zeroed in the background, thereby reducing zeroing time and improving efficiency by optimizing memory access times.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional sequential zeroing methods are used, then memory pages are zeroed in a straightforward manner, but the process becomes time-consuming and hardware intensive as memory sizes grow
Solution Approach 1:
The patent divides the large memory region into multiple smaller sub-regions or pages that can be zeroed independently and in parallel. This segmentation allows the zeroing operation to be distributed across multiple worker threads, each handling a portion of the memory, thereby reducing the overall time required while maintaining complete zeroing coverage
Solution Approach 2:
The patent performs preliminary actions by allocating and initializing memory regions before they are fully needed. By pre-zeroing memory pages in the background using worker threads before the main application requires them, the system reduces the critical path time and avoids blocking the main execution flow
2Productivity
If traditional zeroing processes are used, then the zeroing operation completes the task, but it causes prolonged teardown times and visible delays
Solution Approach 1:
The patent implements periodic action by using background worker threads that perform zeroing operations asynchronously and periodically rather than blocking the main thread. This allows the system to continue other operations while periodically making progress on zeroing memory, improving overall productivity without extending visible teardown time
Solution Approach 2:
The patent introduces worker threads as intermediary components that mediate between the main application thread and the memory zeroing operation. These intermediary threads handle the time-consuming zeroing task in the background, allowing the main thread to proceed without waiting, thereby improving productivity while reducing perceived teardown time
3Speed
If parallel zeroing with worker threads is used, then zeroing speed is improved, but system complexity increases
Solution Approach 1:
The patent applies dynamics by making the number of worker threads configurable and adaptable based on system conditions. Rather than using a fixed number of threads, the system can dynamically adjust thread creation and termination based on available resources and workload, achieving high zeroing speed while managing system complexity through flexible, condition-based thread management
Data Source
AI summary
A method for zeroing memory in computing systems where access to memory is non-uniform includes receiving, via a system call, a request to delete a memory region, forwarding the request to an intermediate software thread, and using the intermediate software thread to perform the request as a background process. The method further includes, upon receiving a message from the intermediate software thread, returning to a system caller, while performing the request, via the intermediate software thread, continues in the background.