EE380 Computer Systems Colloquium

"Build one and throw it out," so the adage goes. This talk explores the argument for systems architecture revision, including processors, operating systems, and networking--both as a general principle and the ways the Internet in particular is currently in need of a reboot. Assumptions, resources, and goals change with time and experience, and so too does our understanding of architectural principles. Through the eyes of the Internet and other examples, we review what we got right (one ring to rule them all, good enough rather than perfect), what we got wrong (7 layers, name resolution as afterthought), and what we only now are beginning to appreciate (layering and forwarding as one thing). Challenge cases are presented that can help drive this redesign, including single-packet exchanges and recursive layering, and an example given of one direction this approach can lead. Finally, we explore the challenges of evolution and transition to help us prepare for giving the Internet a well-deserved reset.

REMOTE URL SU-EE380-20200311

The increase in velocity, volume, and complexity of the data generated by the upcoming Linac Coherent Light Source upgrade (LCLS-II) at the SLAC National Accelerator Laboratory presents a considerable challenge for data acquisition, data processing, and data management. These systems face formidable challenges due to the extremely high data throughput, hundreds of GB/s to multi-TB/s, generated by the detectors at the experimental facilities and to the intensive computational demand for data processing and scientific interpretation. The LCLS-II Data System is a fast, powerful, and flexible architecture that includes a feature extraction layer designed to reduce the data volumes by at least one order of magnitude while preserving the science content of the data. Innovative architectures are required to implement this reduction with a configurable approach that can adapt to the multiple science areas served by LCLS. In order to increase the likelihood of experiment success and improve the quality of recorded data, a real-time analysis framework provides visualization and graphically-configurable analysis of a selectable subset of the data on the timescale of seconds. A fast feedback layer offers dedicated processing resources to the running experiment in order to provide experimenters feedback about the quality of acquired data within minutes. We will present an overview of the LCLS-II Data System architecture with an emphasis on the Data Reduction Pipeline (DRP) and online monitoring framework.

While our software systems have become increasingly elastic, the physical substrate available to run that software (that is, the computer!) has remained stuck in a bygone era of PC architecture. Hyperscale infrastructure providers have long since figured this out, building machines that are fit to purpose -- but those advances have been denied to the mass market. In this talk, we will talk about our vision for a new, rack-scale, server-side machine -- and how we anticipate advances like open firmware, RISC-V, and Rust will play a central role in realizing that vision.

[speaker photo] Sylvain's research at fast.ai has focused on designing and improving techniques that allow models to train fast with limited resources. He has also been a core developer of the fastai library, including implementing the warping transformations, the preprocessing pipeline, much of fastai.text, and a lot more.
Prior to fastai, Sylvain was a Mathematics and Computer Science teacher in Paris for seven years. He taught CPGE, the 2-year French program that prepares students for graduate programs at France's top engineering schools (the "grandes écoles"). After relocating to the USA in 2015, Sylvain wrote ten textbooks covering the entire CPGE curriculum. Sylvain is an alumni from École Normale Supérieure (Paris, France) and has a Master's Degree in Mathematics from University Paris XI (Orsay, France). He lives in Brooklyn with his husband and two sons.fastai is a deep learning library which provides practitioners with high-level components that can quickly and easily provide state-of-the-art results in standard deep learning domains, and provides researchers with low-level components that can be mixed and matched to build new approaches. It aims to do both things without substantial compromises in ease of use, flexibility, or performance. This is possible thanks to a carefully layered architecture, which expresses common underlying patterns of many deep learning and data processing techniques in terms of decoupled abstractions. These abstractions can be expressed concisely and clearly by leveraging the dynamism of the underlying Python language and the flexibility of the PyTorch library. fastai includes: a new type dispatch system for Python along with a semantic type hierarchy for tensors; a GPU-optimized computer vision library which can be extended in pure Python; an optimizer which refactors out the common functionality of modern optimizers into two basic pieces, allowing optimization algorithms to be implemented in 4-5 lines of code; a novel 2-way callback system that can access any part of the data, model, or optimizer and change it at any point during training; a new data block API; and much more. We have used this library to successfully create a complete deep learning course, which we were able to write more quickly than using previous approaches, and the code was more clear. The library is already in wide use in research, industry, and teaching.

This talk will explore Centaur's microprocessor-design technology includes both a new high-performance x86 core AND the industry's FIRST integrated AI Coprocessor for x86 systems. The x86 microprocessor cores deliver high instructions/clock (IPC) for server-class applications and support the latest x86 extensions such as AVX 512 and new instructions for fast transfer of AI data. The AI Coprocessor is a clean-sheet processor designed to deliver high performance and efficiency on deep-learning applications, freeing up the x86 cores for general-purpose computing. Both the x86 and AI Coprocessor technologies have now been proven in silicon using a new scalable SoC platform with eight x86 cores and an AI Coprocessor able to compute 20 trillion AI operations/sec with 20 terabytes/sec memory bandwidth. This SoC architecture requires less than 195mm2 in TSMC 16nm and provides an extensible platform with 44 PCIe lanes and 4 channels of PC3200 DDR4.

Observation tools for understanding occasionally-slow performance in large-scale distributed transaction systems are not keeping up with the complexity of the environment. The same applies to large database systems, to real-time control systems in cars and airplanes, and to operating system design.

Extremely low-overhead tracing can reveal the true execution and non-execution (waiting) dynamics of such software, running in situ with live traffic. KUtrace is such a tool, based on small Linux kernel patches recording and timestamping every transition between kernel- and user-mode execution across all CPUs of a datacenter or vehicle computer. The resulting displays show exactly what each transaction is doing every nanosecond, and hence shows why unpredictable ones are slow, all with tracing overhead well under 1%. Recent additions to KUtrace also show interference between programs and show profiles within long execution stretches that have no transitions.

The net result is deep insight into the dynamics of complex software, leading to often-simple changes to improve performance.

There are dozens of companies selling products to solve your Identity and Access Management (IAM) problems. Why so many? Because IAM is really hard. But why is IAM so hard? I hope to convince you in this talk the reason is that our approach to IAM is fundamentally flawed. The talk includes my personal theory of how we got into this mess and a strategy for getting out of it.

Over the past four years, a federal appeals court has upended a basic assumption about software: that while software can be covered by copyright law, software interfaces cannot. The court ruled that Google's re-implementation of Java APIs infringed Oracle's Java copyrights. Now, the Supreme Court will weigh in. The Electronic Frontier Foundation has argued at each stage of the case that copyright in APIs threatens innovation, especially by smaller players. This is part of EFF's work promoting the freedom to tinker, innovate, and research, especially outside the walls of major corporations—work that has also included challenges to overbroad computer crime and "anti-circumvention" laws.

In the field of Artificial Life, we use biochemistry, engineering and computer simulation to answer "what is life" and "what life could be". Starting from Conway's Game of Life, a famous discrete cell simulation, we tried to make everything smooth and continuous, and then something strange emerged. In this new system called "Lenia", we discovered lots of self-organizing, self-regulating dynamic patterns, look and behave much like microscopic life seen in biology class or plankton videos. More than just being amused, we investigated these creatures in details, performed computer experiments, classification, statistical analysis, and probed into the nature of such dynamic and complex system. We see possible links between Lenia, biological life, and artificial intelligence.

YouTube's recommendation algorithm is frequently characterized by journalists and researchers as radicalizing users to the far-right, but the evidence to date has been weak. We used data collected from the YouTube website to analyze the balance in recommendation impressions to see if it is favoring more extreme content. 768 US political channels were categorized into culturally relevant orientations and sub-cultures and 23M recommendations for recent videos were recorded during November-December 2019. We found that the late 2019 recommendation algorithm actively discourages viewers from being presented with fringe content. The algorithm is shown to favor mainstream media and cable news content over independent YouTube channels with a slant towards partisan political channels like Fox News and Last Week Tonight