Rowhammer Cache: A Last-Level Cache for Low-Overhead Rowhammer Tracking

The rowhammer attack on modern DRAM systems is here to stay as the number of row activations required to induce a DRAM bit flip (rowhammer threshold) is following a trend of concern: 100K activations in 2014 to a few hundred activations in recent years. Hardware mitigations of rowhammer attacks need a rowhammer tracker that can track the DRAM row activations, and trigger the rowhammer mitigation. The state-of-the-art rowhammer tracker named Hybrid Tracker (HYDRA) is a lightweight hardware approach. HYDRA incurs a performance slowdown of less than 1%. However, our evaluations show that HYDRA fails to deliver its promise in terms of performance and storage overhead. For SPEC CPU2017 homogeneous 8-core workloads, HYDRA incurs a performance slowdown of 23.63%. We find that the simulation infrastructure used in HYDRA does not simulate a modern processor with a detailed cache hierarchy with hardware prefetching. The workloads used do not capture all the regions of interest (sim-points) of a benchmark. To mitigate this problem of performance slowdown without compromising security, we propose rowhammer cache, a storage-efficient and low-performance overhead HYDRA that provides the sweet spot in terms of storage overhead, and performance overhead. rowhammer cache uses existing last-level cache (LLC) space instead of additional SRAM storage for rowhammer tracking. Rowhammer cache improves the effectiveness of HYDRA by improving the performance slowdown from 23.6% to 2.25% for 55 8-core SPEC CPU2017 homogeneous mixes. For 20 8-core GAP mixes, it improves the performance slowdown from 36.47% to 5.61%. For 45 representative heterogeneous mixes, the Rowhammer cache improves performance slowdown from 28.68% to 3.14%. Rowhammer cache provides these performance improvements with negligible storage of 512 bytes.