DAIET

Data Aggregation In nETwork


Problem

Network communications are often a scalability bottleneck for partition/aggregate applications

In applications like MapReduce, machine learning, graph processing, data and computations are distributed among many servers and partial results exchanged over the network may generate a large number of flows that overrun buffers and receivers

Approach

DAIET performs data aggregation along network paths using programmable network devices

Offloading to switches and smart NICs part of the computation reduces traffic on the datacenter fabric and the amount of work at destination end hosts

This alleviates communication bottlenecks and can improve overall job completion times

Motivating Scenarios

Experimental results show that the updates sent at the end of an iteration in Machine learning algorithms have a significant overlap (35%-68%), which represents the potential gain of in-network aggregation

Similarly, aggregation of messages exchanged between vertices in Graph Analytics algorithms can provide a large traffic reduction (48%-93%)

Motivating Scenarios

DAIET deployment

Members

KAUST

Publications

Paper

Amedeo Sapio, Ibrahim Abdelaziz, Abdulla Aldilaijan, Marco Canini and Panos Kalnis.
In-Network Computation is a Dumb Idea Whose Time Has Come.
In 16th ACM Workshop on Hot Topics in Networks (HotNets 2017), Palo Alto, California, USA.
PDF [Bibtex]

@inproceedings{daiet,
  title={In-Network Computation is a Dumb Idea Whose Time Has Come},
  author={Sapio, Amedeo and Abdelaziz, Ibrahim and Aldilaijan, Abdulla and Canini, Marco and Kalnis, Panos},
  booktitle={Proceedings of the Sixteenth ACM Workshop on Hot Topics in Networks},
  year={2017}
}

Poster

Amedeo Sapio, Ibrahim Abdelaziz, Marco Canini and Panos Kalnis.
DAIET: A System for Data Aggregation Inside the Network.
In ACM Symposium on Cloud Computing 2017, Santa Clara, California, USA.

@inproceedings{daiet_poster,
  title={DAIET: A System for Data Aggregation Inside the Network},
  author={Sapio, Amedeo and Abdelaziz, Ibrahim and Canini, Marco and Kalnis, Panos},
  booktitle={Proceedings of the Eight ACM Symposium on Cloud Computing},
  year={2017}
}

Code

Our source code, documentation and experiments are on Github.

Resources

Our state-of-the-art testbed comprises an Edgecore Wedge100BF-65X Switch with the cutting edge Barefoot Tofino 6.5 Tbps switch chip and 18 high-end server machines.

Edgecore Wedge100BF-65X
Testbed