Ravan Nazaraliyev

Applied Fault Tree Analysis (FTA), a reliability method long used in aerospace and nuclear power, to root-cause diagnosis of hardware errors in Google's multi-node production servers, with the goal of reducing mean time to repair (MTTR).

• Extracted the PCIe topology from an existing structural model of the machine and generated fault trees from that topology.
• Built a diagnosis tool that takes an error symptom reported by telemetry, traverses the fault tree and recommends which field-replaceable unit (FRU) to swap.
• Evaluated the tool against the current heuristic, rule-based diagnosis practice, aiming for fewer misattributed root causes.

Hardware security and computer architecture, advised by Prof. Nael Abu-Ghazaleh. More on each project on the Research page.

• MIG isolation on the NVIDIA GH200 (PAGE404, IEEE S&P 2027): showed that hardware-assisted page migration in GPUs can be abused to break isolation between co-located workloads.
• Rowhammer mitigations as attack vectors (Not so Refreshing, USENIX Security 2025): showed that the DRAM Refresh Management (RFM) Rowhammer mitigation can be exploited to attack GPUs.
• Multi-GPU interconnect (NVBleed, ACM CCS 2026; SEED 2024): co-developed contention-based covert and side-channel attacks on the NVIDIA NVLink interconnect.
• Efficient Rowhammer defense (PRACtical, under submission): designed subarray-level counter updates and bank-level recovery isolation to reduce the performance overhead of PRAC, evaluated with the Ramulator DRAM simulator.
• Responsibly disclosed findings to the affected hardware vendors and released attack code on GitHub.

• Led weekly discussion sections.
• Held office hours.
• Supervised laboratory sessions.