Searcharxiv⌕ Search

arXiv subjects

Dong Kim

Publications and source records attributed to Dong Kim.

3 recordsLinked to original sources

VS-Flash: Harnessing Vector Search with Computational Memory-Semantic Flash

As generative AI continues to evolve, retrieval-augmented generation (RAG) systems are driving rapid growth in vector database capacity requirements. As datasets outgrow host DRAM, vector databases rely on MMap-based serving over SSDs to extend capacity. However, high-fan-out ANN searches generate concurrent irregular accesses that contend for host memory and storage resources, degrading bulk throughput and inflating tail latency under mixed workloads. In this paper, we address these challenges through memory-semantic capacity and workload-aware compute placement. VS-Flash exposes flash as a large-capacity memory-semantic space and dynamically selects the execution engine according to runtime query and workload characteristics. It offloads throughput-oriented large-batch queries to a Computational Memory-Semantic Flash (CMS-Flash) device, while executing latency-sensitive small-batch queries on the host CPU. Both paths access the same resident segments, avoiding a separate host-DRAM replica. Across three billion-scale datasets, VS-Flash improves bulk-search throughput over vanilla Milvus by 36.8X on average, with up to a 57.4X improvement. Under mixed workloads, VS-Flash reduces the p99 latency of latency-sensitive requests by up to 24.8X while sustaining device-side bulk search. These results show that VS-Flash delivers scalable vector search beyond host DRAM capacity while jointly improving bulk throughput and small-request tail latency.

cs.DB↗

TempoKV: Timely Staging of LLM KV Caches for Memory-Semantic Flash

Reusable prefix key-value (KV) caches can outgrow GPU memory in large language model (LLM) serving. A memory-semantic flash hierarchy offers SSD-backed capacity with a limited fast tier, but a logical KV hit is not necessarily ready for GPU retrieval. Demand staging exposes SSD latency, whereas immediate staging can reserve fast-tier capacity long before retrieval begins. We present TempoKV, a timing-aware resource-commitment layer that separates early knowledge of reuse from the acquisition of staging resources. It records reusable-KV hits as metadata-only claims and requests commitment when the runtime-estimated time until retrieval falls to the storage-estimated time needed to make KV resident and protected against eviction. These estimates adapt to runtime progress and staging state, while commitment remains subject to available protected capacity. We implement TempoKV in vLLM and LMCache on an SSD-backed CXL memory device without changing request scheduling. Across two models and three prefix cache ratios, TempoKV reduces protected fast-tier byte-time per request by 63-91% versus immediate staging while retaining much of the serving benefit of advance staging. In a fast-tier capacity sweep, output throughput and p95 time to first token (TTFT) remain nearly unchanged as capacity decreases from 100 to 25 GiB. Compared with unmodified LMCache's Device-DAX L1 configuration, TempoKV reduces p95 TTFT by up to 48.0% and increases output throughput by up to 27.8%.

cs.DC↗

A High-Performance Solid-State Disk with Double-Data-Rate NAND Flash Memory

We propose a novel solid-state disk (SSD) architecture that utilizes a double-data-rate synchronous NAND flash interface for improving read and write performance. Unlike the conventional design, the data transfer rate in the proposed design is doubled in harmony with synchronous signaling. The new architecture does not require any extra pins with respect to the conventional architecture, thereby guaranteeing backward compatibility. For performance evaluation, we simulated various SSD designs that adopt the proposed architecture and measured their performance in terms of read/write bandwidths and energy consumption. Both NAND flash cell types, namely single-level cells (SLCs) and multi-level cells (MLCs), were considered. In the experiments using SLC-type NAND flash chips, the read and write speeds of the proposed architecture were 1.65-2.76 times and 1.09-2.45 times faster than those of the conventional architecture, respectively. Similar improvements were observed for the MLC-based architectures tested. It was particularly effective to combine the proposed architecture with the way-interleaving technique that multiplexes the data channel between the controller and each flash chip. For a reasonably high degree of way interleaving, the read/write performance and the energy consumption of our approach were notably better than those of the conventional design.

cs.AR↗