HPC, Big Data and Data Science
Scalable, Automated ML Model Monitoring with KFServing and Hopsworks
In this session, we will present an open-source stream processing architecture, based on Spark Structured Streaming, for automating model monitoring with some experiment results. We use Kafka to log model predictions, KFServing for model serving and a Kubernetes operator for the deployment and configuration of the different components. As for the analysis of inference data, we implemented an extendable monitoring framework on top of Spark Structured Streaming to detect outliers and data drift.
In recent years, MLOps has emerged to bring DevOps processes to the machine learning (ML) development process, aiming at more automation in the execution of repetitive tasks and at smoother interoperability between tools. Among the different stages in the ML lifecycle, model monitoring involves the continuous supervision of the model performance over time, involving the combination of techniques in four categories: outlier detection, data drift detection, explainability and adversarial attacks.
Nowadays, most of the available model monitoring tools follow a scheduled batch processing approach or analyse model performance using isolated subsets of the inference data. However, for the continuous monitoring of models, stream processing platforms show several advantages, including support for continuous data analytics, scalable processing of large amounts of data and first-class support for window-based aggregations useful for concept drift detection.
Additional information
| Type | devroom |
|---|
More sessions
| 2/6/21 |
With the end of Moore’s law, improving single-core processor performance can be extremely difficult to do in an energy-efficient manner. One alternative is to rethink conventional processor design methodologies and propose innovative ideas to unlock additional performance and efficiency. In an attempt to overcome these difficulties, we propose a compiler-informed non-speculative out-of-order commit processor, that attacks the limitations of in-order commit in current out-of-order cores to ...
|
| 2/6/21 |
The Container ecosystem spans from spawning a process into an isolated and constrained region of the kernel at bottom layer, building and distributing images just above to discussions on how to schedule a fleet of containers around the world at the very top. While the top layers get all the attention and buzz, this session will base-line the audiences' understanding of how to execute containers.
|
| 2/6/21 |
Over the years, there has been extensive efforts to improve Apache Spark SQL performance. This talk will introduce the new Adaptive Query Execution (AQE) framework and how it can automatically improve user query performance. AQE leverages query runtime statistics to dynamically guide Spark's execution as queries run along. The talk will go over the main features in AQE and provide examples on how it can improve on the previous static query plans. Finally, we'll present the significant ...
|
| 2/6/21 |
This talk introduces DiscoPoP, a tool which identifies parallelization opportunities in sequential programs and suggests programmers how to parallelize them using OpenMP. The tool first identifies computational units which, in our terminology, are the atoms of parallelization. Then, it profiles memory accesses inside the source code to detect data dependencies. Mapping dependencies to CUs, we create a data structure which we call the program execution tree (PET). Further, DiscoPoP inspects the ...
|
| 2/6/21 |
In this talk we explore two programming models for GPU accelerated computing in a Fortran application: OpenMP with target directives and CUDA. We use an example application Riemann problem, a common problem in fluid dynamics, as our testing ground. This example application is implemented in GenASiS, a code being developed for astrophysics simulations. While OpenMP and CUDA are supported on the Summit supercomputer, its successor, an exascale supercomputer Frontier, will support OpenMP and ...
|
| 2/6/21 |
The European Environment for Scientific Software Installations (EESSI, pronounced as “easy”) is a collaboration between different HPC sites and industry partners, with the common goal to set up a shared repository of scientific software installations that can be used on a variety of systems, regardless of which flavor/version of Linux distribution or processor architecture is used, or whether it is a full-size HPC cluster, a cloud environment or a personal workstation. The EESSI codebase ...
|
| 2/6/21 |
XALT is a tool run on clusters to find out what programs and libraries are run. XALT uses the environment variable LD_PRELOAD to attach a shared library to execute code before and after main(). This means that the XALT shared library is a developer on every program run under linux. This shared library is part of every program run. This talk will discuss the various lessons about routine names and memory usage. Adding XALT to track container usage presents new issues because of what shared ...
|
