Global STAC Live, Fall 2020

• Wes McKinney, Director, Ursa Labs

Apache Arrow is a cross-language development platform for in-memory columnar data processing. Among other things, it contains a language-independent columnar data format designed for fast transport and processing on modern hardware. Now in the top 15 most active Apache projects, Arrow has over 500 contributors and over two million weekly downloads. With the Python and R communities using it to facilitate CPU and GPU computations, as well as uptake by major cloud data warehouses, Arrow has a shot (pun intended) at becoming a universal “native” format for data-intensive computing. As a co-creator of Arrow and a key member of its Project Management Committee, Wes will provide a progress update with particular focus on enhancements that can benefit high-throughput data services, data exchange in distributed systems, and C++ analytics on time series, both in-memory and out-of-core. Come to get up to date and ask Wes your questions.

• Pawel Szopinski, VP, Specialised Infrastructure Engineering, Barclays

The Linux Foundation's Data Plane Development Kit (DPDK) is a set of libraries and network interface drivers to accelerate packet processing workloads running with a wide variety of network cards and CPU architectures. In Pawel's opinion, once DPDK is well understood, it can become indispensable tool. In this talk, he will walk us through the core components and architecture of DPDK, point out some helpful open-source implementations, and discuss potential use cases. Whether you are new to DPDK or are looking to take better advantage of it, Pawel's talk is for you.

• Michel Debiche, Director of Analytics Research, STAC

Michel will present the latest benchmark results from technology stacks for compute-intensive simulations.

• Dr John Ashley, General Manager, Financial Services and Technology, NVIDIA

William Gibson famously said that "the future is already here--it's just not evenly distributed". In John's view, that maxim applies to trading desks as much as any other part of life. Some desks continue doing what has worked for the last decade, hand-crafting algorithms on a relatively narrow slice of market data in, say, Excel or C++. Another group is expanding their data sources and exploiting the benefits of the Python ecosystem, including ML/AI tools. John will argue that the future belongs to the latter. He will explain how Python makes it easy for traders to create smarter strategies, and he'll offer tips on how to get smarter faster. By smarter, he means taking in more contextual data and exploiting the latest ML techniques, which he'll demonstrate by showing how a BERT-class model can turn natural language into insights. By faster, John means both faster to market and faster in the market. To accelerate strategy development and backtesting, he will share some best practices and Python community resources. To get ML models to react more quickly to markets, John will discuss the importance of optimizing inference and how to do it. Come to see this view of the future and ask your questions about it.

• Alexander Tsyplikhin, Senior AI Engineer, Graphcore

AI is an increasingly important competitive differentiator for financial firms. But according to Alex, the traditional constraints of current hardware architectures often force firms to oversimplify models in order to reduce latency and fight noise. These compromises are especially limiting as model sizes continue to grow. In Alex's view, if hardware constraints can be relaxed sufficiently, then AI practitioners will be able to use new techniques in order to develop more advanced models with higher accuracy. In this talk, Alex will detail some of these new approaches to machine intelligence and explain how more complex financial models can enable deeper insights faster.

• Michel Debiche, Director of Analytics Research, STAC

Many of the performance challenges associated with machine learning (ML) are in the training phase: that is, model development. But for certain use cases like time-sensitive trading, the inference phase (the application of a model to new data) also has its challenges. For example, if a model provides realtime input to trading algorithms--or perhaps even does the trading itself--inference latency may be a prime consideration. And if the model is deployed in a resource-constrained environment like a colocation center, throughput per unit of resource is probably also a concern. In this talk, Michel will present proposed STAC benchmarks for realtime inference, using an example with long short-term memory (LSTM) networks. He will argue that benchmarks of inference latency will not only inform sizing decisions for well-established ML techniques but will also be a key factor in deciding whether less proven ML techniques can or cannot be applied effectively to various financial use cases.

• Rob Otter, MD & CTO, Applied Research & API Technologies, JP Morgan Chase
• Emmanuel Hauptmann, Founding Partner, RAM Active Investments
• Oleg Rasskazov, Executive Director, London Machine Learning Centre, JP Morgan Chase
• Shahzad Chohan, Global Head of Machine Intelligence & Accelerated Computing, Credit Suisse

Imagine the scenario: A group within your firm wants to explore the use of a particular machine learning technique for a given problem, with an intent to take anything successful into production. Your job is to supply them with the platform(s) they should use for both training and inference. How should you go about this? The number of AI platforms is growing steadily (CPUs, GPUs, FPGAs, vector engines, specialized processors, bare metal/cloud…). Given that different platforms are probably good for different tasks, what is the right way to frame the problem? Are we getting to a point where every new modeling challenge requires a customized hardware evaluation? What role can industry benchmarks play? Beyond speeds and feeds, what are the key business factors to keep in mind? Do small firms have an advantage or can large firms act nimbly as well? To what extent do cloud offerings make the problem easier or harder? Our panel of experts will discuss.

• Anurag Khandelwal, Assistant Professor of Computer Science, Yale University

"Serverless" has become a sexy term that vendors use in many different ways. However, the original innovation that inspired the label is functions as a service (FaaS). FaaS breaks from the paradigm of the constantly-running application on a known virtual or physical server. Instead, developers design applications as discrete functions, each of which is implemented in a cloud framework that invokes a function in response to events and only charges for the time that the function executes. All the major cloud providers offer FaaS today, and at previous STAC summits some user firms have said they cut their spend by 90% on major applications by moving to FaaS. In this talk, Anurag will review the kinds of workloads currently suited to FaaS, as well as the limitations that make it unusable for other workloads--in particular, stateful apps. To do this, he will discuss the backend services required to make FaaS work today and argue that new backend services now in development will make FaaS viable for stateful apps and even streaming analytics. If you need to think about your application architecture a couple of years out, this talk should give you plenty to consider.