Variable Execution-Time Atomic Operations with Adaptive Cache Locking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing memory device designs struggle to efficiently handle both short and long execution time atomic operations (AMOs) in a single memory side cache, leading to performance issues due to the difficulty in accommodating varying run-times of these operations.
Innovation Solution
Implementing a cache system with a results queue and credit-based mechanism to manage long AMOs by locking cache lines during execution and using a results queue, while allowing short AMOs to execute without locking, thereby maintaining efficiency for both types of operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single memory side cache is used to handle both short and long AMOs, then the cache structure is simple, but performance deteriorates due to inability to accommodate varying run-times
Solution Approach 1:
The patent segments the cache handling mechanism into two distinct paths: one for short AMOs that can complete within a single cycle, and another for long AMOs that require multiple cycles. This is achieved by evaluating the run length of each AMO and routing it through appropriate execution paths, thereby resolving the performance issue while maintaining reasonable structural complexity
Solution Approach 2:
The patent introduces dynamic handling of AMOs by evaluating their run length at execution time. Short AMOs are executed immediately without locking, while long AMOs trigger cache line locking and result queuing. This dynamic adaptation allows the system to optimize performance based on the specific characteristics of each AMO, resolving the contradiction between simplicity and performance
2Reliability
If cache line locking is applied to all AMOs, then atomicity is ensured, but productivity decreases due to cache line contention
Solution Approach 1:
The patent applies cache line locking selectively rather than uniformly to all AMOs. Specifically, locking is applied only to long AMOs that require multiple cycles to complete, while short AMOs execute without locking. This localized application of the locking mechanism maintains atomicity where needed while minimizing contention, thereby resolving the contradiction between reliability and productivity
3Productivity
If short AMOs are allowed to execute without locking, then productivity increases, but reliability may be compromised for long AMOs
Solution Approach 1:
The patent dynamically determines the execution strategy for each AMO based on its run length evaluation. Short AMOs are executed without locking to maximize throughput, while long AMOs trigger the locking protocol to ensure atomicity. This dynamic decision-making process allows the system to optimize for productivity when safe and ensure reliability when necessary, resolving the contradiction between these two parameters
Data Source
AI summary
System and techniques for variable execution time atomic operations are described herein. When an atomic operation for a memory device is received, the run length of the operation is measured. If the run length is beyond a threshold, a cache line for the operation is locked while the operation runs. A result of the operation is queued until it can be written to the cache line. At that point, the cache line is unlocked.