Several trends are changing the underlying fabric for data processing in fundamental ways. On the hardware side, machines are becoming heterogeneous with smart NICs, TPUs, DPUs, etc., but specially with GPUs taking a more dominant role. On the software side, the diversity in workloads, data sources, and data formats has given rise to the notion of composable data processing where the data is processed across a variety of engines and platforms. Finally, on the infrastructure side, different storage types, disaggregated storage, disaggregated memory, networking, and interconnects are all rapidly evolving, which demands a degree of customization to optimize data movement well beyond established techniques. To tackle these challenges, in this paper, we present Maximus, a modular data processing engine that embraces heterogeneity from the ground up. Maximus can run queries on CPUs and GPUs, can split execution between CPUs and GPUs, import and export data in a variety of formats, interact with a wide range of query engines through Substrait, and efficiently manage the execution of complex data processing pipelines. Through the concept of operator-level integration, Maximus can use operators from third-party engines and achieve even better performance with these operators than when they are used with their native engines. The current version of Maximus supports all TPC-H queries on both the GPU and the CPU and optimizes the data movement and kernel execution between them, enabling the overlap of communication and computation to achieve performance comparable to that of the best systems available, but with a far higher degree of completeness and flexibility.

In this study we explore the impact of different combinations of GPU models (RTX3090, A100, H100, GraceHoppers - GH200) and interconnects (PCIe 3.0, PCIe 4.0, PCIe 5.0, and NVLink 4.0) on various relational data analytics workloads (TPC-H, H2O-G, ClickBench). We present MaxBench, a comprehensive framework designed for benchmarking, profiling, and modeling these workloads on GPUs. Beyond delivering detailed performance metrics, MaxBench estimates query execution performance using a novel cost model. With this model, we move beyond traditional metrics such as arithmetic intensity and GFlop/s and suggest using instead the notions of characteristic query complexity and characteristic GPU efficiency , as more suitable metrics for data analytics workloads. We conduct an extensive experimental analysis with MaxBench across different combinations of GPU models and interconnects on various data analytics workloads. The insights from this analysis reveal the trade-offs between GPU computing capacity and interconnect bandwidth on query processing. Using this cost model, we also examine future trends by investigating how enhancements in interconnect bandwidth or GPU efficiency would affect performance in the future.

We propose COSMA: a parallel matrix-matrix multiplication algorithm that is near communication-optimal for all combinations of matrix dimensions, processor counts, and memory sizes. The key idea behind COSMA is to derive an optimal (up to a factor of 0.03% for 10MB of fast memory) sequential schedule and then parallelize it, preserving I/O optimality. To achieve this, we use the red-blue pebble game to precisely model MMM dependencies and derive a constructive and tight sequential and parallel I/O lower bound proofs. Compared to 2D or 3D algorithms, which fix processor decomposition upfront and then map it to the matrix dimensions, it reduces communication volume by up to √ times. COSMA outperforms the established ScaLAPACK, CARMA, and CTF algorithms in all scenarios up to 12.8x (2.2x on average), achieving up to 88% of Piz Daint's peak performance. Our work does not require any hand tuning and is maintained as an open source implementation.