Hierarchische und kompakte Datenrepräsentation von Ereignisspuren auf Petascale-Systemen: Effiziente Größenreduktion zur Laufzeit

Date

Dec 3, 2013

Time

9:00 AM - 10:00 AM

Speaker

Michael Wagner

Affiliation

Institut für Technische Informatik, Professur Rechnerarchitektur

Language

en

Main Topic

Informatik

Other Topics

Informatik

Description

One of the most urgent challenges in event based performance analysis is the enormous amount of collected data. In particular, the number of resulting event trace files is pushing against the limits of today’s parallel file systems. To address this issue, my thesis presents a novel approach to efficiently reduce event trace data and keep it in main memory for the entire measurement runtime. This not only enables measurements of applications on large-scale systems but also eliminates bias caused by intermediate memory buffer flushes. First, the recorded data is compressed with new enhanced encoding techniques that will remarkably increase memory efficiency without introducing additional overhead for the compression. Secondly, strategies for efficient event reduction are applied, which guarantee that data of an event tracing measurement fits into a single memory buffer. However, a traditional single flat memory buffer fails to support efficient event reduction. Thus, the third part is a hierarchical memory buffer approach, which is capable to support event reduction at any grade. Diese Veranstaltung wird unterstützt von Professur für Rechnerarchitektur.

Last modified: Dec 3, 2013, 8:55:42 AM