Research
Modern communication mechanisms, such as Remote Direct Memory Access (RDMA), have significantly lowered the CPU cost of large data transfers. However, these performance advantages can only be leveraged through interleaving the computation with the network communication and through careful management of the RDMA buffers used for sending and receiving data. In this project, we investigate how database operators, like join or aggregation algorithms, need to be designed in order to take full advantage of all the features provided by modern low-latency high-throughput networks such as InfiniBand. Our results show that hardware-concious distributed algorithms exhibit good performance and scalability.
Publications
-
Rack-Scale In-Memory Join Processing using RDMA. SIGMOD 2015 -
Distributed Join Algorithms on Hierarchical Network Topologies. Master Thesis – ETH Zurich
Rack-scale computers are becoming the building blocks of modern data centers. They are characterized by new hardware technologies that promise terrabytes of main memory and thousands of CPU cores connected with high-speed interconnects. These systems face a variety of challenges, one of which is the orchestration of the processing and communication resources in order to achieve maximum performance. In this project, we explore how to design a database for a rack-scale system. We investigate several approaches for distributing a state-of-the-art relational database on a modern InfiniBand cluster and evaluate these techniques in terms of performance, scalability and complexity.
Publications
-
Rack-Scale Data Processing System. Eurosys – WRSC 2015