arXiv subjects

Sang-Won Lee

Publications and source records attributed to Sang-Won Lee.

5 recordsLinked to original sources

arXiv2026

RCC: Speculative Write Versioning with Redo Logs

Modern OLTP engines rely on multi-versioning to eliminate read-write conflicts, yet their concurrency is severely limited for write-write conflicts. The conventional wisdom of updating records in place and immediately causes only one transaction to update a record at a time, and other update-conflicting transactions to wait for the former to commit or abort. Thus, conflicting transactions are serialized. We propose RCC, which leverages redo logs to resolve conflicting writes. A transaction updates a record out of place by creating a speculative write version using the redo log. With multiple speculative versions, RCC allows concurrent transactions to be pipelined after the update-conflicting point. Each update made by a transaction is installed to its record upon commit. This lazy update policy enables lightweight rollback: a transaction aborts by discarding its speculative versions. To realize the performance potential of its speculative versioning, RCC proposes two novel techniques: commit-time deadlock detection and columnar RCC (RCC-C). The former detects cycles only once lazily at commit time and the latter eliminates record-level false WW conflicts by leveraging column-granule redo logs. RCC guarantees serializability using lock-based access-time conflict ordering and dependency graph tracking. We implement RCC on MySQL and PostgreSQL, integrating the concurrency control mechanism with their buffer managers, lock managers, and recovery modules. RCC improves TPC-C's transaction throughput and latency over Vanilla versions by an order of magnitude when running 64 concurrent threads on a machine with 128 cores. RCC-C further boosts throughput and latency by avoiding false conflict-induced deadlocks and unnecessary aborts. For a high-contention YCSB benchmark, commit-time deadlock detection enables RCC to scale to 128 clients while competing schemes do not scale beyond 32 threads.

cs.DB↗

arXiv2025

MatKV: Trading Compute for Flash Storage in LLM Inference

We observe two major trends in LLM-based generative AI: (1) inference is becoming the dominant factor in terms of cost and power consumption, surpassing training, and (2) retrieval augmented generation (RAG) is becoming prevalent. When processing long inputs in RAG, the prefill phase of computing the key-value vectors of input text is energy-intensive and time-consuming even with high-end GPUs. Thus, it is crucial to make the prefill phase in RAG inference efficient. To address this issue, we propose MatKV, a scheme that precomputes the key-value vectors (KVs) of RAG objects (e.g., documents), materializes them in inexpensive but fast and power-efficient flash storage, and reuses them at inference time instead of recomputing the KVs using costly and power-inefficient GPU. Experimental results using Hugging Face's Transformers library across state-of-the-art GPUs and flash memory SSDs confirm that, compared to full KV computation on GPUs, MatKV reduces both inference time and power consumption by half for RAG workloads, without severely impacting accuracy in the question-answering task. Furthermore, we demonstrate that MatKV enables additional optimizations in two ways. First, a GPU can decode text while simultaneously loading the materialized KVs for the next instance, reducing load latency. Second, since decoding speed is less sensitive to GPU performance than KV computation, low-end GPUs can be leveraged for decoding without significantly compromising speed once the materialized KVs are loaded into GPU memory. These findings underscore MatKV's potential to make large-scale generative AI applications more cost-effective, power-efficient, and accessible across a wider range of tasks and hardware environments.

cs.DC↗

arXiv2022

Enlightening Flash Storage to Stream Writes by Objects

For a write request, today flash storage cannot distinguish the logical object it comes from. In such object-oblivious flash devices, concurrent writes from different objects are simply packed in their arrival order to flash memory blocks; hence objects with different lifetimes are multiplexed onto the same flash blocks. This multiplexing incurs write amplification, worsening the performance. Tackling the multiplexing problem, we propose a novel interface for flash storage, FlashAlloc. It is used to pass the logical address ranges of logical objects to the flash storage and thus enlighten the storage to stream writes by objects. The object-aware flash storage can de-multiplex writes from different objects with distinct deathtimes into per-object dedicated flash blocks. Given that popular data stores separate writes using objects (e.g., SSTables in RocksDB), we can achieve, unlike the existing solutions, transparent write streaming just by calling FlashAlloc upon object creation. Our experimental results using an open-source SSD prototype demonstrate that FlashAlloc can reduce write amplification factor (WAF) in RocksDB, F2FS, and MySQL by 1.5, 2.5, and 0.3, respectively and thus improve throughput by 2x, 1.8x, and 1.2x, respectively. In particular, FlashAlloc will mitigate the interference among multitenants. When RocksDB and MySQL were run together on the same SSD, FlashAlloc decreased WAF from 4.2 to 2.5 and doubled their throughputs.

cs.DB↗

arXiv2012

Flash-based Extended Cache for Higher Throughput and Faster Recovery

Considering the current price gap between disk and flash memory drives, for applications dealing with large scale data, it will be economically more sensible to use flash memory drives to supplement disk drives rather than to replace them. This paper presents FaCE, which is a new low-overhead caching strategy that uses flash memory as an extension to the DRAM buffer. FaCE aims at improving the transaction throughput as well as shortening the recovery time from a system failure. To achieve the goals, we propose two novel algorithms for flash cache management, namely, Multi-Version FIFO replacement and Group Second Chance. One striking result from FaCE is that using a small flash memory drive as a caching device could deliver even higher throughput than using a large flash memory drive to store the entire database tables. This was possible due to flash write optimization as well as disk access reduction obtained by the FaCE caching methods. In addition, FaCE takes advantage of the non-volatility of flash memory to fully support database recovery by extending the scope of a persistent database to include the data pages stored in the flash cache. We have implemented FaCE in the PostgreSQL open source database server and demonstrated its effectiveness for TPC-C benchmarks.

cs.DB↗

arXiv2011

B+-tree Index Optimization by Exploiting Internal Parallelism of Flash-based Solid State Drives

Previous research addressed the potential problems of the hard-disk oriented design of DBMSs of flashSSDs. In this paper, we focus on exploiting potential benefits of flashSSDs. First, we examine the internal parallelism issues of flashSSDs by conducting benchmarks to various flashSSDs. Then, we suggest algorithm-design principles in order to best benefit from the internal parallelism. We present a new I/O request concept, called psync I/O that can exploit the internal parallelism of flashSSDs in a single process. Based on these ideas, we introduce B+-tree optimization methods in order to utilize internal parallelism. By integrating the results of these methods, we present a B+-tree variant, PIO B-tree. We confirmed that each optimization method substantially enhances the index performance. Consequently, PIO B-tree enhanced B+-tree's insert performance by a factor of up to 16.3, while improving point-search performance by a factor of 1.2. The range search of PIO B-tree was up to 5 times faster than that of the B+-tree. Moreover, PIO B-tree outperformed other flash-aware indexes in various synthetic workloads. We also confirmed that PIO B-tree outperforms B+-tree in index traces collected inside the Postgresql DBMS with TPC-C benchmark.

cs.DB↗