Thostrup, Lasse ; Failing, Daniel ; Ziegler, Tobias ; Binnig, Carsten (2022)
A DBMS-centric Evaluation of BlueField DPUs on Fast Networks.
48th International Conference on Very Large Databases. Sydney, Australia (05.09.2022-09.09.2022)
Konferenzveröffentlichung, Bibliographie
Kurzbeschreibung (Abstract)
Modern networks have evolved significantly in the last years. First, network speed has increased considerably and thus the use of lowoverhead techniques such as RDMA has become more and more important to design efficient distributed DBMSs. Second, a recent trend in modern networks is that in addition to high-speed data transfer using RDMA, network components such as switches and NICs become programmable by providing additional computation on the device, such as DPUs (Data Processing Units). Such devices enable processing or manipulation of data as it is traversing the network and that way allow distributed systems to offload computation. While for the recent generation of RDMA-based DPU cards, there is no study that shows the offloading capabilities of DBMS tasks to such RDMA-enabled DPUs. Therefore, in this paper, we aim to provide a first systematic study to evaluate the basic performance characteristics of the BlueField network cards in the context of typical DBMS operations. For the evaluation, we analyze the offload potential of using BlueField as a RDMA-enabled DPU for two important use cases: (1) a remote Btree and (2) an end-host sequencer (i.e., remote counter). We chose these two use cases since they represent core tasks where RDMA has shown benefits. As a result, in our evaluation, we show that the recent generation of RDMA-based Bluefield DPUs can provide several benefits and can not only reduce access latencies but also improve the throughput. However, offloading computation to the DPU needs a careful design and naively offloading all computation to the DPU often leads to performance degradation.
| Typ des Eintrags: | Konferenzveröffentlichung |
|---|---|
| Erschienen: | 2022 |
| Autor(en): | Thostrup, Lasse ; Failing, Daniel ; Ziegler, Tobias ; Binnig, Carsten |
| Art des Eintrags: | Bibliographie |
| Titel: | A DBMS-centric Evaluation of BlueField DPUs on Fast Networks |
| Sprache: | Englisch |
| Publikationsjahr: | 9 Mai 2022 |
| Veranstaltungstitel: | 48th International Conference on Very Large Databases |
| Veranstaltungsort: | Sydney, Australia |
| Veranstaltungsdatum: | 05.09.2022-09.09.2022 |
| URL / URN: | https://www.adms-conf.org/adms_2022.html |
| Kurzbeschreibung (Abstract): | Modern networks have evolved significantly in the last years. First, network speed has increased considerably and thus the use of lowoverhead techniques such as RDMA has become more and more important to design efficient distributed DBMSs. Second, a recent trend in modern networks is that in addition to high-speed data transfer using RDMA, network components such as switches and NICs become programmable by providing additional computation on the device, such as DPUs (Data Processing Units). Such devices enable processing or manipulation of data as it is traversing the network and that way allow distributed systems to offload computation. While for the recent generation of RDMA-based DPU cards, there is no study that shows the offloading capabilities of DBMS tasks to such RDMA-enabled DPUs. Therefore, in this paper, we aim to provide a first systematic study to evaluate the basic performance characteristics of the BlueField network cards in the context of typical DBMS operations. For the evaluation, we analyze the offload potential of using BlueField as a RDMA-enabled DPU for two important use cases: (1) a remote Btree and (2) an end-host sequencer (i.e., remote counter). We chose these two use cases since they represent core tasks where RDMA has shown benefits. As a result, in our evaluation, we show that the recent generation of RDMA-based Bluefield DPUs can provide several benefits and can not only reduce access latencies but also improve the throughput. However, offloading computation to the DPU needs a careful design and naively offloading all computation to the DPU often leads to performance degradation. |
| Freie Schlagworte: | systems_funding_52300543, systems_dfg_spp2037, systems_maki, systems_funding_52115350 |
| Zusätzliche Informationen: | 13th International Workshop on Accelerating Analytics and Data Management Systems Using Modern Processor and Storage Architectures |
| Fachbereich(e)/-gebiet(e): | 20 Fachbereich Informatik 20 Fachbereich Informatik > Data and AI Systems |
| TU-Projekte: | DFG|SFB1053|SFB1053 TPZ Steinmet DFG|BI2011/1-1|Skalierbares Datenma |
| Hinterlegungsdatum: | 05 Apr 2023 13:15 |
| Letzte Änderung: | 24 Jul 2023 10:05 |
| PPN: | 509882854 |
| Export: | |
| Suche nach Titel in: | TUfind oder in Google |
 |
Frage zum Eintrag |
Optionen (nur für Redakteure)
 |
Redaktionelle Details anzeigen |
Drucken
Impressum