In this paper, we present a novel communication scheme called zero-sided RDMA, enabling data exchange as a native network service using a programmable switch. In contrast to one- or two-sided RDMA, in zero-sided RDMA, neither the sender nor the receiver is actively involved in data exchange. Zero-sided RDMA thus enables efficient RDMA-based data shuffling between heterogeneous hardware devices in a disaggregated setup without the need to implement a complete RDMA stack on each heterogeneous device or the need for a CPU that is co-located with the accelerator to coordinate the data transfer. As such, we think that zero-sided RDMA is a major building block to make efficient use of heterogeneous accelerators in future cloud DBMSs. In our evaluation, we show that zero-sided RDMA can outperform existing one-sided RDMA-based schemes for accelerator-to-accelerator communication and thus speed up typical distributed database operations such as joins.

In this paper, we present a novel communication scheme called zero-sided RDMA, enabling data exchange as a native network service using a programmable switch. In contrast to one- or two-sided RDMA, in zero-sided RDMA, neither the sender nor the receiver is actively involved in data exchange. Zero-sided RDMA thus enables efficient RDMA-based data shuffling between heterogeneous hardware devices in a disaggregated setup. In our initial evaluation, we show that zero-sided RDMA can outperform existing one-sided RDMA-based schemes due to offloading the coordination to the network and new optimizations that are only possible by coordinating the data exchange on the switch.

In this paper, we present a novel pipelined GPU join that accelerates the performance of distributed DBMSs by leveraging GPU resources on fast networks. A key insight is that we enable pipelined join execution by overlapping the network shuffling with the build and probe phases, thereby significantly reducing the GPU idle time. To demonstrate this, we propose novel algorithms for distributed pipelined GPU joins with RDMA and GPUDirect for both arbitrarily large probe- and build-side tables. In our evaluation, we show our pipelined distributed GPU join can reduce the overall runtime of a full query by up to 6× against a state-of-the-art CPU-only join.

Modern cloud networks are a fundamental pillar of data-intensive applications. They provide high-speed transaction (packet) rates and low overhead, enabling, for instance, truly scalable database designs. These networks, however, are fundamentally different from conventional ones. Arguably, the two key discerning technologies are RDMA and programmable network devices. Today, these technologies are not niche technologies anymore and are widely deployed across all major cloud vendors. The question is thus not if but how a new breed of data-intensive applications can benefit from modern networks, given the perceived difficulty in using and programming them. This tutorial addresses these challenges by exposing how the underlying principles changed as the network evolved and by presenting the new system design opportunities they opened. In the process, we also discuss several hard-earned lessons accumulated by making the transition first-hand.

In this paper we present a new approach for distributed DBMSs called P4DB, that uses a programmable switch to accelerate OLTP workloads. The main idea of P4DB is that it implements a transaction processing engine on top of a P4-programmable switch. The switch can thus act as an accelerator in the network, especially when it is used to store and process hot (contended) tuples on the switch. In our experiments, we show that P4DB hence provides significant benefits compared to traditional DBMS architectures and can achieve a speedup of up to 8x.

Programmable switches allow to offload specific processing tasks into the network and promise multi-Tbit/s throughput. One major goal when moving computation to the network is typically to reduce the volume of network traffic, and thus improve the overall performance. In this manner, programmable switches are increasingly used, both in research as well as in industry applications, for various scenarios, including statistics gathering, in-network consensus protocols, and more. However, the currently available programmable switches suffer from several practical limitations. One important restriction is the limited amount of available memory, making them unsuitable for stateful operations such as Hash Joins in distributed databases. In previous work, an FPGA-based In-Network Hash Join accelerator was presented, initially using DDR-DRAM to hold the state. In a later iteration, the hash table was moved to on-chip HBM-DRAM to improve the performance even further. However, while very fast, the size of the joins in this setup was limited by the relatively small amount of available HBM. In this work, we heterogeneously combine DDR-DRAM and HBM memories to support both larger joins and benefit from the far faster and more parallel HBM accesses. In this manner, we are able to improve the performance by a factor of 3x compared to the previous HBM-based work. We also introduce additional configuration parameters, supporting a more flexible adaptation of the underlying hardware architecture to the different join operations required by a concrete use-case.

In this paper, we propose the Data Flow Interface (DFI) as a way to make it easier for data processing systems to exploit high-speed networks without the need to deal with the complexity of RDMA. By lifting the level of abstraction, DFI factors out much of the complexity of network communication and makes it easier for developers to declaratively express how data should be efficiently routed to accomplish a given distributed data processing task. As we show in our experiments, DFI is able to support a wide variety of data-centric applications with high performance at a low complexity for the applications.