Liuwei Fu

Liuwei Fu

Auto-generated profile, based on publicly available information

Liuwei Fu

School of Software, Dalian University of Technology

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
6 August 2025 in ACM Transactions on Architecture and Code Optimization
Verified
Journal Article
Open access
HJHe Jiang
He Jiang
LFLiuwei Fu
Liuwei Fu
DLDong Liu
Dong Liu
ZRZhilei Ren
Zhilei Ren
YCYuting Chen
Yuting Chen
LQLei Qiao
Lei Qiao
In modern microarchitectures, machine-learning-based prefetchers use past memory requests to learn access patterns and predict memory addresses, thereby prefetching data into the cache to mitigate the processor-memory speed gap. However, they face two key challenges in capturing irregular access patterns generated by complex data structures and algorithms. One is data dispersion: the disorderliness of memory addresses makes it difficult for prefetchers to extract meaningful data features. The other is temporal and spatial complexity: existing prefetchers fail to effectively learn temporal and spatial characteristics, and thus are unable to explore more complex access patterns. To resolve these challenges, we propose TRACED, a novel temporal graph neural network-based prefetcher aimed at learning access patterns of memory addresses. TRACED consists of two key components: a dynamic clustering component and a temporal graph neural network component. In the dynamic clustering component, we introduce a similarity function to quantify the similarity of memory addresses. Based on the quantified similarity, we dynamically group unordered memory addresses into different clusters. This ensures that the memory addresses in each cluster are ordered and change smoothly, thus resolving the first challenge. The temporal graph neural network component constructs a spatiotemporal graph to represent relationships among memory addresses. This helps capture temporal and spatial characteristics both across and within clusters, thus resolving the second challenge. This paper demonstrates the effectiveness of the proposed prefetcher through experiments. Specifically, in terms of accuracy, TRACED outperforms BO, SPP, DOMINO, Delta-LSTM, and VOYAGER by 2.29%-40.83% on average. Furthermore, TRACED attains remarkable coverage of 55.67% and IPC of 43.75%, outperforming all competing approaches in both metrics.

29 August 2026 in ACM Transactions on Architecture and Code Optimization
Verified
Journal Article
Open access
LFLiuwei Fu
Liuwei Fu
HJHe Jiang
He Jiang
AJAng Jia
Ang Jia
PZPeiyu Zou
Peiyu Zou
ZRZhilei Ren
Zhilei Ren
LQLei Qiao
Lei Qiao
Data prefetching techniques have been widely adopted to resolve the bottleneck of access latency between processors and memory. However, existing prefetchers, while effective in hiding memory access latency, often aggressively prefetch large amounts of overpredictions into caches in advance. This behavior can result in cache pollution, bandwidth waste, and even degrade overall system performance. To resolve this issue, numerous filters have been integrated into prefetchers to eliminate overpredictions during prefetching. Nevertheless, existing mainstream filters in prefetchers typically suffer from two critical limitations. First, they lack the awareness to system performance , neglecting the overall impact of prefetch decisions on the system. Second, they exhibit insufficient adaptivity to program behaviors , making them incapable of dynamically adjusting to changes in different programs or execution stages. To overcome these limitations, we propose RELOAF, a RE inforcement L earning-based O nline A daptive F iltering framework that can be integrated into different prefetchers. Specifically, RELOAF comprises two key components: a system information feedback component and an adaptive reinforcement learning (RL) decision component. In system information feedback component, system-level feedback information, such as prefetch accuracy, coverage, and IPC, is introduced as a decision criterion to evaluate the validity of prefetching, thereby resolving the first limitation. In the adaptive RL decision component, RELOAF dynamically adjusts prefetching based on the current program state using reinforcement learning, thus handling the second limitation. Experiments validate the effectiveness of the proposed approach. Results on benchmark workloads show that augmenting state-of-the-art prefetchers with RELOAF improves their average prefetch accuracy by 3.65%–6.49% and IPC by 0.81%–1.83%, compared to the same prefetchers operating without RELOAF.