<  Back to the Polytechnique Montréal portal

Enhanced state history tree (eSHT): a stateful data structure for analysis of highly parallel system traces

Loïc Prieur-Drevon, Raphaël Beamonte, Naser Ezzati-Jivan and Michel Dagenais

Paper (2016)

Open Access document in PolyPublie
[img]
Preview
Open Access to the full text of this document
Accepted Version
Terms of Use: All rights reserved
Download (413kB)
Show abstract
Hide abstract

Abstract

Behaviors of distributed systems with many cores and/or many threads are difficult to understand. This is why dynamic analysis tools such as tracers are useful to collect run-time data and help programmers debug and optimize complex programs. However, manual trace analysis on very large traces with billions of events can be a difficult problem which automated trace visualizers and analyzers aim to solve. Trace analysis and visualization software needs fast access to data which it cannot achieve by searching through the entire trace for every query. A number of solutions have adopted stateful analysis which rearranges events into a more query friendly structures after a single pass through the trace. In this paper, we look into current implementations and model the behavior of previous work, the State History Tree (SHT), on traces with many thread creation and deletion. This allows us to identify which properties of the SHT are responsible for inefficient disk usage and high memory consumption. We then propose a more efficient data structure, the enhanced State History Tree (eSHT), to store and query computed states, in order to limit disk usage and reduce the query time for any state. Next, we compare the use of SHT and eSHT on traces with many attributes. We finally verify the scalability of our new data structure according to trace size. As shown by our results, the proposed solution makes near optimal use of disk space, reduces the algorithm's memory usage logarithmically for previously problematic cases, and speeds up queries on traces with many attributes by an order of magnitude. The proposed solution builds upon our previous work, enabling it to easily scale up to traces containing a million threads.

Uncontrolled Keywords

state system, tracing, parallel, distributed, data structure, stateful analysis, history tree, multi-dimensional indexes

Subjects: 2700 Information technology > 2705 Software and development
2700 Information technology > 2713 Algorithms
2700 Information technology > 2720 Computer systems software
Department: Department of Computer Engineering and Software Engineering
Funders: CRSNG/NSERC
Grant number: CRDPJ468687-14
PolyPublie URL: https://publications.polymtl.ca/2994/
Conference Title: IEEE International Congress on Big Data (BigData Congress 2016)
Conference Location: San Francisco, CA, USA
Conference Date(s): 2016-06-27 - 2016-07-02
Publisher: IEEE
DOI: 10.1109/bigdatacongress.2016.19
Official URL: https://doi.org/10.1109/bigdatacongress.2016.19
Date Deposited: 13 Feb 2018 12:51
Last Modified: 10 Apr 2024 05:51
Cite in APA 7: Prieur-Drevon, L., Beamonte, R., Ezzati-Jivan, N., & Dagenais, M. (2016, June). Enhanced state history tree (eSHT): a stateful data structure for analysis of highly parallel system traces [Paper]. IEEE International Congress on Big Data (BigData Congress 2016), San Francisco, CA, USA (8 pages). https://doi.org/10.1109/bigdatacongress.2016.19

Statistics

Total downloads

Downloads per month in the last year

Origin of downloads

Dimensions

Repository Staff Only

View Item View Item