Near-Data Data Transformation

Note: See the Relational Memory Controller project page for research that builds on this project.

BU faculty members Manos Athanassoulis and Renato Mancuso will work with Red Hat researchers Uli Drepper and Ahmed Sanaullah to create a hardware-software co-design paradigm for data systems that implements near-memory processing. The approach has the potential to revolutionize data management by bridging the gap of analytical and transactional processing. This paradigm addresses the performance bottleneck caused by memory bandwidth and will allow both cloud and edge systems to efficiently handle mixed transactional and analytics data-intensive workloads with a better trade-off between bandwidth and latency. “The proposed software-hardware co-design methodology will also improve the collective understanding of new design models and resource management strategies that are possible in systems with programmable memory hierarchies,” the team wrote.

Summary of Progress:
We built the first implementation of a Relational Memory Engine, that is able to natively access both rows and columns without pushing unnecessary data through the memory hierarchy. The approach is described in an EDBT paper (to be presented in 2023). Further, we have initiated the development of infrastructure building for a configurable memory controller and for an arbitrary data transformation unit. Both approaches will be further followed up in subsequent projects (See the Relational Memory Controller project page).

The research support provided by Red Hat was determinant in completing the first prototype implementation of a data reorganization engine logically located between processors and memory. The engine, namely the Relational Memory Engine, is capable of performing on-the-flight transformation of relational data from the way it is stored in main memory (as a list of data rows) into a configurable layout as seen by the processor(s) and the CPU-side cache hierarchy. The switch in layout allows software logic to be simplified and to achieve better usage of cache resources while attaining better spatiotemporal data locality—and thus performance. The support from Red Hat was not only crucial to seeding the initial idea of the project but also to producing the first proof-of-concept implementation on an ARM aarch64 commercial system-on-a-chip. The work was published at the 26th International Conference on Extending Database Technology (EDBT’23) and will be presented in March 2023. Encouraged by the promising results, we have significantly expanded the scope of the project and applied for additional funding to bring the work to the next level. The team pursued additional funding opportunities with both Red Hat and the National Science Foundation (NSF).

Please see the arxiv version of our ongoing work.

Related Labs

Learn more about the work of Renato and his students at Cyber Physical Systems Lab @ BU

Learn more about the work of Manos and his students at DiSC lab

This work has also received support (a gift) from Cisco

Looking ahead (details to be added)

Accepted paper for publication for EDBT 2023. Title of the paper: “Relational Memory: Native In-Memory Accesses on Rows and Columns“

Presentations

(to be delivered in 2023) At the EDBT 2023 conference “Relational Memory: Native In-Memory Accesses on Rows and Columns”
(to be delivered in 2023) At Microsoft Research, by Manos Athanassoulis (invited talk)
Oct 2022, HPTS 2022, “Transparent Data Transformation” by Manos Athanassoulis (slides: http://hpts.ws/papers/2022/RelMem-HPTS.pdf)
Aug 2022, DevConf 2022 ” Open Hardware Initiative Series: Relational Memory: Native In-Memory Stride Access” by JuHyoung Mun (https://devconfus2022.sched.com/event/14TET/open-hardware-initiative-series-relational-memory-native-in-memory-stride-access)
June 2022, Seminar Talk @ Technical University of Munich (TUM), “Fine-grained Resource Profiling and Knowledge-driven Allocation” by Renato Mancuso
June 2022, Seminar Talk @ University of Minho, “Fine-grained Resource Profiling and Knowledge-driven Allocation” by Renato Mancuso
June 2022, CAPITAL Workshop 2022 (Keynote), “From Memory Partitioning to Management through Fine-grained Profiling and Control” by Renato

Watch