What the study found
Thoth, a hardware prefetcher, was reported to uncover data-dependent memory access (DDMA) patterns more robustly than prior approaches by working with explicit producer-consumer load pairs. The authors say it can handle multi-level range relations that earlier methods struggled to learn.
Why the authors say this matters
The authors argue that DDMA patterns are common in sparse data structures used in graph analytics, machine learning, and high-performance computing, and that missed prefetching opportunities can hurt memory performance. They conclude that Thoth helps address those limitations by reducing mismatches in sampled load instances and improving pattern discovery.
What the researchers tested
The researchers presented Thoth, which detects producer-consumer load pairs using register-level dependency tracking and then uses annotation-directed load sampling to sample only annotated load instances. They also used precise load annotation with reorder identifiers to handle pipeline flushes.
What worked and what didn't
On a suite of DDMA-intensive benchmarks, Thoth achieved a 51.1% speedup over a no-prefetching baseline. The abstract says it outperformed two state-of-the-art DDMA prefetchers by 14.7% and 8.2%, respectively, while prior address-based and instruction-based methods struggled with multi-level range relations, miscorrelation, mismatched load instances, or incomplete dependency chains.
What to keep in mind
The available summary does not describe benchmark details beyond stating that they were DDMA-intensive. It also does not provide limitations, overheads, or information about how the method performs outside the reported benchmark suite.
Key points
- Thoth is a hardware prefetcher for data-dependent memory access patterns.
- The method uses explicit producer-consumer load pairs instead of dependency chains.
- Annotation-directed load sampling is used to sample only matched load instances.
- On DDMA-intensive benchmarks, Thoth showed a 51.1% speedup over no prefetching.
- Thoth outperformed two state-of-the-art DDMA prefetchers by 14.7% and 8.2%.
Disclosure
- Research title:
- Thoth improves prefetching for irregular memory access patterns
- Authors:
- Kanheng Jiang, Yongxin Lyu, Z Y Zhang, Zengshi Wang, Chao Fu, Jun Han
- Institutions:
- Fudan University, Fudan University, Fudan University, Fudan University, Fudan University, Fudan University, Shanxi University
- Publication date:
- 2026-04-03
- DOI:
- W7148931474
- OpenAlex record:
- View
Get the weekly research newsletter
Stay current with scholarly research without reading academic papers — one filtered digest, every Friday.