In a rapidly evolving digital landscape, managing large data lakes and streaming data presents

unique challenges.

Balancing storage, accessibility, and real-time analysis of these colossal data sets is crucial

for businesses to unlock their true potential and drive innovation.

Processing large-scale data consumes a lot of energy.

Reducing the power consumption while maintaining efficiency is vital for data centers.

This balance is key to sustainable, powerful data management.

Efficient energy use doesn't only cut costs, but also reduces the environmental impact.

In the classic three-layer architecture, where data processing is separate from data storage,

there's a significant increase in data movement.

This separation often leads to higher energy consumption, as vast amounts of data continuously

shuttle back and forth between storage and servers.

By moving processing closer to storage, we drastically cut down the data transferred

over the network, leading to more efficient and eco-friendly operations.

Based on this principle, we've developed Reprovide, a heterogeneous data processing

system.

This system is uniquely capable of handling both large data sets and real-time data, leveraging

the power of FPGA-based programmable system on chips.

Reprovide stands as a testament to our innovative approach, showcasing the seamless integration

of hardware and software in processing complex data efficiently.

Our system integrates an x86-based host with a uniquely designed cluster of FPGA-based

PSOCs.

Each RPU boasts a powerful Zilinx FPGA and an ARM CPU complemented by multiple 10Gbit

Ethernet connections.

Additionally, each RPU is directly connected to several high-speed NVMe storage devices.

Each FPGA within our system is composed of multiple reconfigurable regions, versatile

in their functionality.

To better visualize this concept, we use LEGO building blocks.

Each block represents an operation on the data stream.

These regions are capable of housing different accelerators, designed to perform a variety

of operations, including, among others, projecting and filtering.

The accelerator is not static.

It's synthesized either beforehand or dynamically, on the fly, as and when required.

The RPUs are housed in a 19-inch rack, at the heart of which lies a 10Gbit switch, dedicated

to the cluster network.

This design not only ensures efficient connectivity, but also allows seamless scalability.

To unleash the full potential of our system, we need an optimizer that considers all the

limitations and potentials our system has.

The goal is to maximize throughput and minimize latency while maintaining a low level of energy

consumption.

To solve this problem, we first need to address some challenges.

A key challenge our system faces lies in optimizing queries to suit our specific hardware architecture,

particularly in managing the limited resources available on our RPUs.

This requires strategic resource allocation to maximize efficiency and performance within

the constraints of our hardware setup.

To tackle this challenge, we've developed a new optimizer named Kraken.

Kraken combines a novel cost model meticulously crafted for modern hardware, and a resource

model that specifically caters to our unique requirements.

This innovative approach ensures that Kraken efficiently manages and optimizes resources,

aligning perfectly with the demands of our system.