Introduction
As DRAM technology scales to smaller and denser nodes, memory cells become increasingly susceptible to RowHammer (RH). This security vulnerability occurs when rapid activations of a specific memory row cause electrical interference, leading to bit-flips in adjacent rows. To mitigate this, modern memory controllers employ hardware trackers to monitor row activations and trigger proactive refreshes of potential victim rows.
However, as the RowHammer threshold ($N_{RH}$) drops—from 70,000 to below 1,000 activations—tracking every row individually has become impractical due to storage constraints. This has led to the development of "scalable" trackers that share resources, such as counters or cache entries. This paper introduces DAPPER (Performance-Attack-Resilient Tracker), a defense mechanism designed to thwart a new class of threats: Performance Attacks (Perf-Attacks).
The Problem: Performance Attacks on Shared Trackers
While scalable trackers like Hydra, START, CoMeT, and ABACUS successfully prevent bit-flips, their shared architectural structures introduce a deterministic side-channel. An adversary can exploit these structures by crafting specific memory access patterns that overwhelm the tracker.
As shown in Figure 3, these attacks typically fall into two categories:
- Inducing Misses: Attackers can cause frequent misses in shared counter caches (e.g., Hydra's RCC or START's reserved LLC region), forcing slow DRAM accesses for counter updates.
- Triggering Repetitive Refreshes: Attackers can manipulate shared counters (e.g., CoMeT's RAT or ABACUS's spillover counters) to trigger constant, redundant mitigative refreshes, starving benign applications of memory bandwidth.
These attacks can degrade the performance of co-running benign applications by up to 90%, a far more severe impact than standard cache thrashing.
Proposed Solution: DAPPER
DAPPER addresses these vulnerabilities through two design iterations:
DAPPER-S (Simple)
DAPPER-S utilizes a secure, low-latency block cipher (LLBC) to randomly map row addresses to shared counters. This randomization prevents attackers from targeting specific counters. However, it remains partially vulnerable to "streaming" attacks where an attacker touches many rows to overflow all counters.
DAPPER-H (High-resilience)
DAPPER-H, the flagship design, introduces three innovations to achieve near-perfect resilience:
- Double-Hashing: Rows are mapped to two independent counter tables. A refresh is only triggered if both counters reach the threshold, exponentially lowering the success rate of an attacker.
- Per-Bank Bit-Vectors: A bit-vector tracks if a row has been accessed within a refresh window. If the bit is not set, the tracker updates the bit but ignores the shared counter, effectively filtering out malicious streaming patterns.
- Selective Reset: Instead of resetting all counters simultaneously—a predictable event attackers can exploit—DAPPER-H resets only the specific counters involved in a mitigation.
Evaluation
The authors evaluated DAPPER using 57 workloads from major benchmark suites (SPEC2006, SPEC2017, TPC, Hadoop, MediaBench, and YCSB) simulated on a dual-channel DDR5-6400 system.
As illustrated in Figure 1, existing trackers suffer massive performance hits under attack, with normalized performance dropping as low as 0.1 in some cases. In contrast, DAPPER-H maintains a near-baseline performance (represented by the 1.0 line), incurring only a 0.9% average slowdown even at an ultra-low $N_{RH}$ of 500.
Key Results
- Resource Efficiency: DAPPER-H requires only 96KB of SRAM per 32GB of DRAM, significantly less than proposals requiring large associative memories or megabytes of storage.
- Energy Impact: The design is energy-efficient (~0.1% overhead) as it eliminates the redundant refreshes that plague other trackers under stress.
Conclusion
DAPPER demonstrates that RowHammer defense must consider not only security against bit-flips but also resilience against performance sabotage. By combining secure hashing with intelligent filtering, DAPPER provides a scalable, low-overhead solution that remains robust even as DRAM density continues to increase and RowHammer thresholds continue to fall.
Summary updated