Disaggregated Memory Systems (DMS) hold substantial potential for cloud datacenters but face critical deployment barriers in multi-tenant RDMA environments. Existing DMS designs rely on idealized assumptions-overlooking interference from co-located RDMA applications, oversimplifying fabric topology considerations, and lacking elastic service-level objectives (SLOs) guarantees-resulting in performance degradation and resource inefficiency.

The Log-Structured Merge-tree (LSM-tree) has become a cornerstone of modern key-value stores (KVSs) due to its efficiency in handling write-intensive workloads. However, traditional monolithic LSM-tree designs suffer from write stalls caused by resource contention between Memtable flushing and SSTable compaction, while existing distributed systems adopt coarse-grained elasticity that limits resource utilization and responsiveness. This paper introduces ServerlessLSM, a kernelspace RDMA-odesigned Serverless workflow architecture for LSM-trees. By decoupling Memtable flushing and compaction into independent Serverless functions, ServerlessLSM enables fine-grained elasticity and low-latency state transfers through distributed OS primitives (e.g., remote fork, remote memory mapping). Evaluations demonstrate that ServerlessLSM reduces cold-start latency by$\mathbf{9 8 \%}$and achieves$\mathbf{2. 4} \times$higher throughput compared to state-of-the-art solutions, while maintaining space amplification below 11 %, validating its efficiency and costeffectiveness in cloud environments.

Large Language Model (LLM) inference suffers from inefficiencies in coupled prefill (P) and decoding (D) phases, leading to resource underutilization and scheduling bottlenecks. While disaggregated P-D architectures address this by isolating phases across asymmetric clusters, serverless deployments introduce critical challenges: cold-start latency during autoscaling and costly intermediate state transfers (e.g., KV cache) between distributed prefill and decoding instances. We present ServerlessPD, a system that co-designs remote fork with RDMA to enable near-instant autoscaling and zero-copy state transferring for serverless LLM inference. ServerlessPD introduces a RDMA-based OS kernel-integrated primitive that remotely forks active prefill instances into decoding instances across machines, bypassing cold starts by reusing pre-materialized GPU states, which grants child containers direct copy-on-write access to parent GPU memory. The system further employs GPU context interception to efficiently capture and replicate execution states, ensuring seamless state transfer. To optimize resource utilization, ServerlessPD integrates a dynamic launch-point algorithm that schedules fork operations based on real-time prefill-decoding dynamics, minimizing idle time and overlapping computation with state transfers. ServerlessPD demonstrates that RDMA-codeigned remote fork can unlock near-instant autoscaling and efficient state disaggregation for LLM serving.

Persistent key-value store based on LSM-trees represents one of the most advanced designs. Recent research shows that key-value separation has become a popular optimization method for LSM-tree. However, this storage architecture still incurs significant overhead when dealing with some query- and update-intensive workloads. In this paper, we propose$\text{LC}\text{KV}$, a key-value separated LSM-tree storage system built using the$\underline{L}earner-\underline{C}leaner$optimization to increase throughput. Learner represents the construction of learned indexes to increase query throughput, responsible for building models for the hot-readcold-written keys stored in the LSM-tree and values stored in the cold-written$\mathrm{v}\text{alue logs}$(vLogs). Cleaner represents the garbage$\mathrm{c}\text{ollector}$(GC) aimedat increasing update throughput, responsible not only for garbage collection but also for maintaining the sorting of the cold-written vLog. Evaluations show that LCKV outperforms other state-of-the-art solutions.