Processing-in-Memory (PIM) is a paradigm promising to reduce data movement as a growing bottleneck in data-intensive systems such as database systems. The main idea of PIM is to bring computation to memory, requiring to rethink architectures and designs of data management solutions. In this tutorial, we give an overview of the application of the PIM paradigm to data management tasks using the UPMEM PIM technology.

While near-memory computing offers significant speedups for parallel database workloads, the only currently available commercial implementation, UPMEM PIM, suffers from setup and data transfer overheads.

Due to these, its past applications in database systems have focused on analytical (read-only) workloads with data permanently residing in UPMEM modules, and dynamic partial offloading of queries that takes these overheads into account.

Here, we examine a middle ground: letting data reside in UPMEM modules across consecutive queries, and evicting it only when it is altered by a write statement or when another data set should be offloaded instead.

Specifically, we combine benchmark-based performance models with a task-based scheduling simulator to determine at which point offloading to UPMEM is worth the incurred overheads for a variable number of available CPU cores, UPMEM ranks, table sizes, and consecutive UPMEM-compatible database queries.

We show that this approach leads to reliable latency simulations that can be verified with real-world benchmarks.

Using these simulations, we find that that offloading individual columns to UPMEM becomes worthwhile once table size exceeds 2²² to 2²⁷ rows, with a small number of consecutive queries (one to several dozen) needed to amortize the setup and data transfer overheads.

The size of datasets in DBMSs are growing super-linearly, increasing the demand for storage devices as data can no longer fit entirely in main memory.

Data movement between the CPU and storage devices often limits performance, primarily due to poor data locality, which is difficult to optimize in DBMSs.

An emerging solution to this challenge is the use of Computational Storage Devices (CSDs), which offload certain computation tasks to the storage hardware.

However, due to the high cost of fully programmable CSDs, current solutions often offer limited computation capabilities, such as the Scaleflux CSD-3000, which features transparent compression.

Transparent compression allows for more flexibility in the storage layout of data structures and DBMS storage formats.

As a result, traditional techniques for reducing storage overhead in B$^+$-Trees can be simplified or eliminated, leading to higher bandwidth and reduced storage overhead.

In this work, we evaluate the CSD-3000 with transparent compression by examining commonly used DBMS data structures and storage layouts.

Our experiments cover a variety of DBMSs with different storage formats.

The results demonstrate that transparent compression reduces storage overhead while improving bandwidth through better data locality.