Processing-In-Memory Primitives for Data Management

Processing-In-Memory Primitives for Data Management (PIMPMe) is a joint project of the Computer Architecture and Embedded Systems Group and the Databases and Information Systems Group at Technische Universität Ilmenau. The project is part of the DFG priority programme on Disruptive Memory Technologies (SPP 2377).

Motivation

Traditional computer architectures are characterized by a clear separation of tasks: while processors such as CPU or GPU perform computations, other components such as storage, memory or network only “deliver” data. Thus, data is constantly being transferred between, e.g., memory and processor, and with the rapidly increasing capacity of memory and storage, as well as the growing demands of applications in terms of data size, data transfer has a significant impact on performance.

Preliminary work

As part of the first phase of the priority program, we exhaustively studied the utilization of commercially available Processing-In-Memory technologies, especially the UPMEM DIMMs, for various data management tasks. Therefore, we considered memory-bound database operations in accordance with the architectural peculiarities of UPMEM. Building on our research results, we (re-)implemented query operators and index structures. Overall, the results of our work have demonstrated the performance benefits of PIM for bandwidth-bound and latency-bound workloads, resulting from its high data parallelism and on-chip direct memory access.

Objective for Phase 2

While we have already explored the PIM paradigm as PNM in the first phase of the priority program, another technological development opens up additional perspectives. Emerging interconnect technologies such as Remote Direct Memory Access (RDMA) and especially Compute Express Link (CXL) allow bypassing the CPU for data exchange, even for remote memory. For example, the CXL 3.0 specification provides low-latency, memory-semantic communication, allowing interconnected hosts to communicate peer-to-peer (P2P) and share memory as so-called global fabric attached memory (GFAM). In addition, CXL integrates cache coherency with cache line granularity access over PCIe, promising reduced overhead and complexity. Combining such a GFAM with the PIM paradigm would allow not only memory sharing but also in-memory computation, which is the main subject of this Project.

Timeline