Database Prefetching Framework Speeds Buffer Recovery When Workload Hotspots Shift

Database researchers have unveiled a buffer prefetching framework that tackles a stubborn performance problem: when the data access hotspots in a workload shift dramatically, the buffer hit rate plunges and can take a long time to recover. The new system, called pre-buffer, was described in a study published in the journal Data Science and Engineering on 21 May 2026 by Jie Yi, Xiaoliang Wang and Peiquan Jin of the University of Science and Technology of China. In experiments on two tailored workloads, the framework shortened the time needed for the buffer hit rate to rebound after each hotspot switch, and it cut total execution time compared with several established buffer replacement algorithms.

The problem the authors set out to solve is one they name the buffer cold-start problem. Modern database systems serve workloads in finance, e-commerce, social media and logistics, and their access patterns increasingly break the traditional assumptions of temporal and spatial locality on which classic buffer algorithms rely. Buffer management is a crucial component of database performance because its main task is to reduce disk input and output operations and speed up data access. The most common replacement algorithms, such as LRU, LFU and CLOCK, depend on empirical assumptions about locality. Under stable workloads they maintain a high buffer hit rate, but when the workload changes drastically they struggle to evict old hot data in time, so new hot data can enter only passively through page faults. The result is a sharp drop in the hit rate and a slow recovery.

The authors point to evidence that hotspots really do churn quickly in production systems. Measurements of Facebook’s production workloads show that more than two-thirds of the objects in the top 10 percent most requested during one hour drop out of that group in the following hour, and at a finer granularity about half of the popular objects lose their popularity within ten minutes. Other studies have explicitly identified hotspot shift as a core challenge and proposed detection mechanisms to adapt to hotspot migration. Adaptive replacement algorithms such as ARC, which dynamically balances the recency and frequency of data, improve buffer management to a degree, but the authors argue that when hot and cold data swap completely, such algorithms still cannot actively predict which data will become hot next, leaving the hit rate to recover slowly. Newer replacement strategies, including SIEVE and S3-FIFO, have also been proposed, yet few address the cold-start issue effectively.

The key insight behind pre-buffer is that many real-world database workloads are periodic. Access patterns fluctuate regularly over time, with common cycles including intraday peaks on e-commerce platforms, intensive transaction processing in financial systems at the end of the month, and complex analytical queries in data warehouses during quarterly report generation. If hotspots recur periodically, the system can collect hotspot information the first time it appears and then prefetch the same pages when a similar hotspot shows up again. Pre-buffer therefore divides the buffer into a main buffer, which responds to user read and write requests using an LRU replacement strategy, and a separate prefetch buffer that stores pages loaded proactively by an independent prefetching thread. When a user reads a page from the prefetch buffer, a memory tuner module migrates it into the main buffer.

The workflow relies on what the authors call workload awareness, split into two stages. In initial awareness, the system analyzes data access hotspots over a period and records a new hotspot dataset because no similar hotspot has been observed before. In secondary awareness, the system detects that the current hotspot resembles one collected previously and prefetches the corresponding pages into the prefetch buffer in advance. Detection is triggered by a monitor thread that counts the buffer hit rate every second: when the hit rate in one second falls more than a predefined constant, SHIFT_THRESHOLD, below the previous second, the system judges that a hotspot has shifted. A tracker then temporarily collects page access information, and the prefetcher computes the similarity between this fresh access sequence and every hot page set collected earlier. If the maximum similarity exceeds a preset threshold, the matching hot page set is loaded into the prefetch buffer; otherwise the system treats the hotspot as new and restarts its collection.

Similarity calculation is the mathematical heart of the framework, and the authors implement and compare two measures. The first is the classic Jaccard Index, which reflects the overlap of two sets through the ratio of their intersection to their union, ignoring how often each element appears. The second is the Weighted Jaccard Index, which incorporates access frequencies as weights: the weight of a page in the test sequence is its access count divided by the sequence length, in a manner similar to term frequency, while pages in a collected hot page set are weighted by their share of total visits. In the actual framework the test access sequence used for comparison is 1,000 pages long. The weighted version proved more robust in testing: as the prefetch buffer size changed, the similarity values for previously seen hotspots stayed concentrated between 0.1 and 0.15, so the threshold did not need repeated adjustment, whereas the unweighted Jaccard Index required experimental retuning whenever the buffer size changed. The weighted measure also produced a larger gap in similarity between segments that should and should not trigger prefetching.

To evaluate the design, the researchers built two workloads based on the Zipfian distribution with a skew parameter of 0.9, following the generation method used in the YCSB benchmark, over a 6 GB database file of 1.5 million pages with a 1 GB main buffer. Workload W1 divided the pages into four parts corresponding to four distinct hotspots, with a switching sequence in which each hotspot appears at least twice, totaling 12 million page accesses. Workload W2 used three randomly shuffled arrays to create hotspots that partially overlap and are harder to distinguish, with a switching sequence of 10.5 million accesses. The SHIFT_THRESHOLD constant was set to 10 percent after statistical analysis of 100 hotspot switching events showed the vast majority of hit rate drops exceeded that level. The baselines were LRU, ARC, S3-FIFO and SIEVE without prefetching, and the default prefetch buffer held 80,000 pages of 4 KB, roughly 320 MB.

The main experiments showed that in the workload segments where prefetching should trigger, the hit rate recovered markedly faster than all baselines, rising steeply after each drop. The authors measured recovery time as the interval needed for the hit rate to climb back to 71 percent, the average stable level of the baselines, in the segments that triggered prefetching, and found the prefetch framework significantly reduced this time on both workloads. They also measured total execution time over five consecutive runs: LRU had the longest runtime, the newer cache algorithms were shorter, and adding the prefetch framework reduced execution time further. Performance under the two similarity measures was very similar, and the heavier computation of the Weighted Jaccard Index did not affect system performance in the experiments.

Sensitivity analyses refined the picture. Varying the prefetch buffer size across 40,000, 60,000, 80,000 and 100,000 pages showed that larger buffers sped recovery, with 80,000 or 100,000 pages allowing most of the recovery to complete within one second of the drop. Small buffers of 40,000 pages offered limited benefit. Because the prefetch buffer effectively expands total buffer capacity, the stabilized hit rate also ended up above the baseline. The authors suggest that with a 1 GB main buffer, a prefetch buffer of 60,000 to 80,000 pages, roughly 20 to 30 percent of the main buffer, is a reasonable setting, and that 40 percent or more can be used when faster recovery is required, subject to memory constraints. Threshold experiments showed the trade-offs clearly: raising the similarity threshold to 0.12 or 0.14 caused prefetching to fail when it should trigger, while a very low threshold of 0.001 led the system to prefetch unwanted pages and to skip collecting genuinely new hotspot information, yielding minimal improvement.

The authors are careful about the scope of their results. The experiments ran on two simulated workloads designed around periodic hotspot recurrence, not on full production database systems, and the framework by design leans on periodic update patterns in the data. In their future work they list several limitations to address: the strategy’s generality under heavier and more diverse access patterns still has room for improvement, the SHIFT_THRESHOLD was determined experimentally and may need adaptive tuning in more complex environments, hotspot collection currently considers only page visit counts rather than dimensions such as access order, and combining prefetching with adaptive buffer replacement strategies remains unexplored. They also note that more sophisticated machine learning pattern recognition would need careful evaluation to ensure its benefits outweigh its performance overhead. Within those boundaries, the study offers a concrete demonstration that recognizing recurring hotspots and prefetching them ahead of demand can sharply shorten the painful cold-start period that database buffers endure every time the workload’s center of gravity moves.

Subject of Research: A workload-aware buffer prefetching framework that mitigates the buffer cold-start problem in database systems under periodic workloads.

Article Title: Workload-Aware Buffer Prefetching for Database Systems

Article References: Yi, J., Wang, X., & Jin, P. (2026). Workload-Aware Buffer Prefetching for Database Systems. Data Science and Engineering. https://doi.org/10.1007/s41019-025-00342-6

Image Credits: AI Generated

DOI: 10.1007/s41019-025-00342-6

Keywords: database systems, buffer management, prefetching, buffer hit rate, workload awareness, hotspot shifts, LRU, Jaccard Index, YCSB, cold-start problem, cache replacement, periodic workloads