[Paper Review] DReAM: Dynamic Re-arrangement of Address Mapping to Improve the Performance of DRAMs

Mohsen Ghasempour, Jim Garside+2 more|arXiv (Cornell University)|Sep 12, 2015

Parallel Computing and Optimization Techniques14 references4 citations

TL;DR

DReAM is a novel hardware technique that dynamically reconfigures DRAM address mapping at runtime based on workload-specific memory access patterns, using entropy-based bit analysis to minimize page conflicts. It improves DRAM performance by up to 28% on average across workloads, outperforming static mapping schemes without requiring application停机 or software changes.

ABSTRACT

The initial location of data in DRAMs is determined and controlled by the 'address-mapping' and even modern memory controllers use a fixed and run-time-agnostic address mapping. On the other hand, the memory access pattern seen at the memory interface level will dynamically change at run-time. This dynamic nature of memory access pattern and the fixed behavior of address mapping process in DRAM controllers, implied by using a fixed address mapping scheme, means that DRAM performance cannot be exploited efficiently. DReAM is a novel hardware technique that can detect a workload-specific address mapping at run-time based on the application access pattern which improves the performance of DRAMs. The experimental results show that DReAM outperforms the best evaluated address mapping on average by 9%, for mapping-sensitive workloads, by 2% for mapping-insensitive workloads, and up to 28% across all the workloads. DReAM can be seen as an insurance policy capable of detecting which scenarios are not well served by the predefined address mapping.

Motivation & Objective

• To address the performance degradation in DRAMs caused by static, fixed address-mapping schemes that do not adapt to dynamic memory access patterns.
• To reduce page conflicts—caused by consecutive accesses to different rows in the same bank—by dynamically reassigning address mappings based on real-time access behavior.
• To enable on-the-fly address mapping reconfiguration without halting applications, overcoming limitations of software-only or offline approaches.
• To design a low-overhead mechanism that supports runtime data migration while minimizing storage and hardware costs.

Proposed method

• DReAM monitors memory access patterns by approximating the entropy of each physical address bit over a sliding window of 250K requests to detect workload-specific access behavior.
• It uses a counter and XOR gate per address bit to track bit transitions and identify high-entropy bits that correlate with access locality.
• Based on the detected access pattern, DReAM computes an optimized address-mapping scheme that minimizes row buffer conflicts and page conflicts.
• It performs online data migration using a migration table (MT) and a state table (ST) to relocate data across banks, avoiding intra-bank moves to reduce overhead.
• The system supports dynamic re-mapping by tracking last-accessed addresses and enabling incremental data movement during normal operation.
• Storage overhead is minimized, with only 60 bytes for monitoring and 3×10⁻⁵% of DRAM size for migration metadata, making it scalable and practical.

Experimental results

Research questions

• RQ1 Can a runtime-adaptive address-mapping scheme significantly reduce page conflicts in DRAMs compared to fixed, static schemes?
• RQ2 What is the performance gain of dynamically reconfiguring address mapping based on real-time access patterns across diverse workloads?
• RQ3 How can online data migration be efficiently and cost-effectively implemented to support runtime address mapping changes without application interruption?
• RQ4 What is the storage and hardware overhead of maintaining runtime address-mapping adaptation and data migration state?
• RQ5 Can the system detect workload-specific access patterns well enough to justify the complexity and overhead of dynamic re-mapping?

Key findings

• DReAM-Offline outperforms the best static address-mapping scheme by 5% on average and up to 28% across all 48 workloads.
• For mapping-sensitive workloads, DReAM achieves an average performance improvement of 9% over the best baseline.
• For mapping-insensitive workloads, DReAM still improves performance by 2% on average, demonstrating broad applicability.
• In the online variant, DReAM-Online improves performance by 4.5% on average for 12 workloads that meet its detection threshold, with a peak gain of 23%.
• The system imposes only 60 bytes of monitoring overhead and 3×10⁻⁵% of DRAM size for migration metadata, indicating minimal storage cost.
• 87.5% of data relocation occurs between banks, and intra-bank relocation is avoided to reduce migration cost, confirming the efficiency of the design.