Bloom Filter Memory Collision Detection for Parallel Threads
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing approaches for memory address collision detection in parallel threads using bloom filters are inefficient due to the need for multiple bloom filters and high propagation costs, especially when dealing with a large number of threads.
Innovation Solution
A single load bloom filter and a single store bloom filter are used to track memory addresses for all parallel threads, with each storage location expanded to store both a bit indicating access and the numerical order of the highest ordered thread that accessed it, allowing for efficient detection of out-of-order operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple bloom filters are used to track memory addresses for parallel threads, then collision detection accuracy is improved, but device complexity and propagation costs increase
Solution Approach 1:
The patent merges multiple bloom filters into a single bloom filter by expanding each storage location to store both a bit indicating access and the numerical order of the highest ordered thread that accessed it. This consolidation reduces device complexity while maintaining collision detection accuracy through the expanded storage structure that tracks thread ordering information.
Solution Approach 2:
The single bloom filter is designed to serve multiple functions: it tracks memory address accesses across all parallel threads, detects collisions, and determines thread ordering relationships. The expanded storage locations simultaneously store access status and thread order information, making the structure multi-functional and eliminating the need for separate filters for each thread.
2Reliability
If multiple bloom filters are used for each parallel thread, then memory address tracking is improved, but propagation costs increase
Solution Approach 1:
The patent combines multiple bloom filters into one unified structure, reducing the propagation cost associated with updating and querying multiple separate filters. The single bloom filter with expanded storage locations processes all thread memory addresses centrally, eliminating the overhead of coordinating multiple filters while maintaining accurate tracking of each thread's memory access pattern.
3Ease of operation
If traditional bloom filters are used, then implementation simplicity is maintained, but ability to detect out-of-order operations is insufficient
Solution Approach 1:
The patent enhances the traditional bloom filter by applying local quality expansion to the storage locations. Each location is expanded to contain both a bit indicating access and a numerical value representing the highest ordered thread that accessed it. This localized enhancement at each storage position enables out-of-order operation detection while keeping the overall bloom filter structure and operation simple.
Data Source
AI summary
A semiconductor chip is described having a load collision detection circuit comprising a first bloom filter circuit. The semiconductor chip has a store collision detection circuit comprising a second bloom filter circuit. The semiconductor chip has one or more processing units capable of executing ordered parallel threads coupled to the load collision detection circuit and the store collision detection circuit. The load collision detection circuit and the store collision detection circuit is to detect younger stores for load operations of said threads and younger loads for store operations of said threads.