Sie haben Javascript deaktiviert!
Sie haben versucht eine Funktion zu nutzen, die nur mit Javascript möglich ist. Um sämtliche Funktionalitäten unserer Internetseite zu nutzen, aktivieren Sie bitte Javascript in Ihrem Browser.

CRC 901 – On-The-Fly Computing (OTF Computing) Show image information

CRC 901 – On-The-Fly Computing (OTF Computing)


Second SFB 901 Seminar in winter semester 2019/2020

Begin: Wed, 20. of November 2019 ( 4:00 PM)
Location: Fürstenallee 11, Room F0.530 + F0.225

On November 20, 2019, the 2nd SFB 901 seminar in the winter semester will take place.

16:00 - 16:20 subproject C2
Speaker: Prof. Dr. Marco Platzner
Title: The Case for Heterogeneous Migration

Heterogeneous computing has become an emerging topic in high-performance as well as embedded computing domains in recent years. Application tasks expose different computational demands that can be best served by different architectures such as CPUs, GPU, and FPGAs in terms of execution time and energy requirement. A key topic of sub-project C2 is to develop and evaluate an execution model for application tasks that allows for runtime-migration across the different architectures of a heterogeneous compute node. In this introductory talk, we will first give a brief overview over this work. Then, we will present HEFT, a state-of-the-art technique for scheduling a set of dependent tasks on heterogeneous resources. The following in-depth talk will show how we can improve on HEFT by exploiting the novel feature of runtime migratability.

16:20 - 16:40 subproject B3
Speaker: Prof. Dr. Heike Wehrheim
Title: Validating Properties of Machine Learning Models

17:00 - 17:40 parallel Session

Room: F0.225
Speaker: Achim Lösch, subproject C2
Title: MigHEFT: DAG-based Scheduling of Migratable Tasks on Heterogeneous Compute Nodes

In this talk we will present an execution model for heterogeneous migration and its practical realization on a high-performance heterogeneous compute node. Leveraging this novel feature of heterogenous task migration, we will focus on the problem of scheduling a set of dependent tasks, specified by a directed acyclic graph (DAG), to a heterogeneous architecture with the objective to minimize the makespan. For this we will show how the baseline HEFT scheduling heuristic can be adapted to our type of heterogeneous environment. Then, we will detail our novel scheduling approach MigHEFT, that creates an off-line schedule based on estimates for the task execution times. At runtime, we execute this schedule and monitor task executions. In case there is a definite deviation from the off-line schedule, we re-schedule the task set, possibly inducing migration of already running tasks. Due to this migration capability, we achieve an average speedup of 16.1% over HEFT.

Room: F0.530
Speaker: Arnab Sharma, subproject B3
Title: Validating Properties of Machine Learning Models

One task of subproject B3 is the quality assurance of software compositions. In the context of AutoML, such compositions often contain machine learning (ML) models. Thus there is a need for developing validation techniques for ML components. This is challenging due to the fact that the behavior of such components is not programmed but learned.

In this talk, we present a testing approach for ML models which targets the validation of one specific property: monotonicity. Our approach generates test inputs for a black-box ML model via verification and counter example generation on a white-box model. Thereby, we can systematically construct test inputs and improve over randomly generated inputs.

The University for the Information Society