We believe that the benefits of new hardware are best understood in the context of an end-to-end application performance. Although microbenchmarks are easier to develop and provide more impressive numbers (by zeroing in on exactly the component of the system that has been sped up), they can be misleading as to the practical impact of hardware innovations on user-observable performance. Hence, we aim to develop end-to-end and robust applications which integrate with the Crossroads server platform; these applications will both illuminate the benefits of the Crossroads 3D-FPGA architecture and also stand on their own as best-of-breed exemplars of software systems which integrate with an FPGA. Over the three-year period of the project, we will work with three end-to-end applications: an intrusion detection system (IDS), a DNN-based sentiment analysis system, and an operating system for microservices. These full benchmarks will better serve as technology drivers for innovation.
Pigasus: Intrusion Detection at 100Gbps Pigasus is a first-of-its-kind open-source intrusion detection system to operate over 100Gbps+ within the footprint of a single server. The key to Pigasus’ success is an FPGA-centric design where the majority of traffic is filtered on a Stratix-10 MX; Pigasus can serve 100Gbps using only 29 CPU cores (where the state-of-the-art software design requires as many as 700 cores) and requiring 25x less power. We will use Pigasus to kick off our benchmarking process; we will explore integrating Pigasus with new overlays (RV2) and use Pigasus as a starting application to experiment with VTR-3D (RV4) early in our development process. Finally, we will integrate components of Pigasus into the Crossroads Modular Shell (NMS) such as PCIe integration, pattern matching, and TCP reassembly.
Rivulet: Streaming, Multi-Source Sentiment Analysis (Capstone) As our capstone application, we are developing a framework for machine learning on-the-move as data is read from network, disk, or memory. As an exemplar, we will perform sentiment analysis over Twitter data (read in from one of the above sources) with results collated on the CPU and reported over the network to an external database. The Rivulet pipeline will use core components of the NMS (pattern matching, TCP reassembly) and integrate an FPGA-accelerated Neural Network; this application will be the core of the center’s capstone demo.
NormanOS Our most ambitious application is a new operating system meant to take advantage of the Crossroads 3D-FPGA `under the hood’, allowing unmodified applications to benefit from FPGA acceleration over their data on-the-move. As a first step, we focus on pushing networking functionality (everything under the socket) to the Crossroads 3D-FPGA; we can do this with an end-to-end prototype during our three-year period using a Stratix-10 MX as our initial platform then moving to Agilex. In the long run (post this three-year initial project), we expect that the NormanOS will push other OS functionality into the Crossroads 3D-FPGA, e.g., a new file system which compresses and encrypts data within the Crossroads 3D-FPGA and en route to storage, batching and formatting inline for data pushed to GPUs within the Crossroads 3D-FPGA, etc. Using NormanOS as an enabling platform, we will also prototype illustrative examples of new in-interconnect capabilities that complement other in-network elements (e.g., Barefoot Tofino/P4) to realize self-driving network capabilities.
RV1 PIs: Justine Sherry and Vyas Sekar