Referee1
- R1-1. Justification of the 80% migration gate despite lower p99 latency at 70%
- R1-2. Clarification of the asynchronous migration execution in Algorithm 1
- R1-3. Consolidation of redundant descriptions in Section 3
- R1-4. Condensation of the DeathStarBench request-path text in Section 2
- R1-5. Ownership and promotion priority for pages shared across co-located services
- R1-6. Handling of multiple concurrent bottleneck services within a single epoch
- R1-7. Clarification of the host-to-CXL tier capacity configuration
Referee2
- R2-1. Effectiveness and path saturation under irregular access patterns
- R2-2. Relationship to proactive prefetching into the fast memory tier
Referee3
- R3-1. Unrealistic deployment assumption (colocated microservices)
(a) Practical relevance of single-node colocation assumption in modern cloud-native deployments.
(b) Observability limits of local Recv-Q under cross-node bottlenecks
- R3-2. Fragility of TCP Recv-Q as a bottleneck signal
- R3-3. Comparison against state-of-the-art kernel-assisted tiering systems
- R3-4. Limited conceptual novelty and system integration
(a) Fundamental conceptual novelty beyond integration of existing components
(b) Thrashing avoidance under simultaneous backlogs with insufficient DRAM
Editors' comments