Region Victim Hash for Coherence Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Region Coherence Arrays face challenges in maintaining inclusion over a processor's cache hierarchy, leading to difficult line evictions that reduce cache hit rates and performance, especially when scaled down in size, as they require large arrays to avoid unnecessary broadcasts and cache evictions.
Innovation Solution
Implementing a Region Victim Hash (RVH) that tracks evicted lines from the Region Coherence Array without requiring eviction from the processor's cache hierarchy, using a non-tagged hash table of counts to maintain inclusion without evicting lines, allowing for precise tracking and reducing the need for large arrays.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If Region Coherence Arrays are scaled down in size to reduce memory usage, then device complexity and memory requirements are reduced, but cache hit rates decrease and performance deteriorates due to increased line evictions
Solution Approach 1:
The patent divides the coherence tracking function into two separate data structures: a compact Region Coherence Array (RCA) for quick lookup and a larger Region Victim Hash (RVH) for detailed tracking. This segmentation allows the RCA to remain small while the RVH handles the bulk of coherence tracking, resolving the contradiction between small array size and high cache hit rate.
Solution Approach 2:
The Region Victim Hash acts as an intermediary between the compact RCA and the cache hierarchy. When a line is evicted from the RCA, the RVH tracks it without requiring eviction from the cache, serving as a mediator that maintains inclusion property while enabling small RCA size.
2Reliability
If lines are evicted from the processor's cache hierarchy to maintain inclusion in Region Coherence Arrays, then inclusion property is maintained, but cache hit rates and performance are reduced
Solution Approach 1:
Instead of physically evicting lines from the cache hierarchy, the patent creates a copy-tracking mechanism in the Region Victim Hash. The RVH maintains a count of evicted lines, allowing the system to track inclusion property without actual cache evictions, thus maintaining both inclusion and performance.
Solution Approach 2:
The patent changes the parameter representation from binary (line present/absent in cache) to a count-based system in the RVH. This allows tracking of multiple evicted lines without actual evictions, maintaining inclusion property while avoiding performance degradation from cache evictions.
3Productivity
If large Region Coherence Arrays are used to avoid line evictions and maintain high cache hit rates, then cache hit rates are improved, but device complexity and memory requirements increase
Solution Approach 1:
The patent segments the coherence tracking into a small RCA for fast access and a larger RVH for comprehensive tracking. This allows the system to achieve high cache hit rates through the RVH's detailed tracking while keeping the RCA small for low memory usage and complexity.
Solution Approach 2:
The patent adds a new dimension to coherence tracking by introducing the RVH as a separate data structure. Instead of enlarging the RCA in one dimension, the solution moves to another dimension by creating a complementary tracking structure, achieving comprehensive tracking without increasing RCA size.
4Loss of energy
If two-level cache coherence protocols broadcast memory requests to local SMP node first, then global bandwidth requirements are reduced, but memory requests are delayed by local checking
Solution Approach 1:
The patent performs preliminary action by pre-tracking evicted lines in the Region Victim Hash before actual cache access occurs. This allows the system to quickly determine whether a line might be cached without immediate broadcasting, reducing both global bandwidth consumption and latency by avoiding unnecessary local checks.
Solution Approach 2:
The RVH provides feedback information about evicted lines that helps determine whether broadcasting is necessary. This feedback mechanism allows the system to make informed decisions about memory request routing, reducing both bandwidth consumption and latency by avoiding unnecessary local SMP node checks.
Data Source
AI summary
A method and system for precisely tracking lines evicted from a region coherence array (RCA) without requiring eviction of the lines from a processor's cache hierarchy. The RCA is a set-associative array which contains region entries consisting of a region address tag, a set of bits for the region coherence state, and a line-count for tracking the number of region lines cached by the processor. Tracking of the RCA is facilitated by a non-tagged hash table of counts represented by a Region Victim Hash (RVH). When a region is evicted from the RCA, and lines from the evicted region still reside in the processor's caches (i.e., the region's line-count is non-zero), the RCA line-count is added to the corresponding RVH count. The RVH count is decremented by the value of the region line count following a subsequent processor cache eviction/invalidation of the region previously evicted from the RCA.