| Kuan-Hsun Chen, Björn Bönninghoff, Jian-Jia Chen and Peter Marwedel.
Compensate or Ignore? Meeting Control Robustness Requirements through Adaptive Soft-Error Handling.
In Languages, Compilers, Tools and Theory for Embedded Systems (LCTES)
Santa Barbara, CA, U.S.A., June 2016
[BibTeX][PDF][Link][Abstract]@inproceedings { Chenlctes2016,
author = {Chen, Kuan-Hsun and B\"onninghoff, Bj\"orn and Chen, Jian-Jia and Marwedel, Peter},
title = {Compensate or Ignore? Meeting Control Robustness Requirements through Adaptive Soft-Error Handling},
booktitle = {Languages, Compilers, Tools and Theory for Embedded Systems (LCTES)},
year = {2016},
address = {Santa Barbara, CA, U.S.A.},
month = {June},
organization = {ACM},
url = {http://dx.doi.org/10.1145/2907950.2907952},
keywords = {kuan},
file = {http://ls12-www.cs.tu-dortmund.de/daes/media/documents/publications/downloads/2016-khchen-lctes.pdf},
confidential = {n},
abstract = {To avoid catastrophic events like unrecoverable system failures on mobile and embedded systems caused by soft-errors, software-
based error detection and compensation techniques have been proposed. Methods like error-correction codes or redundant execution can offer high flexibility and allow for application-specific fault-tolerance selection without the needs of special hardware supports. However, such software-based approaches may lead to system overload due to the execution time overhead. An adaptive deployment of such techniques to meet both application requirements and system constraints is desired. From our case study, we observe that a control task can tolerate limited errors with acceptable performance loss. Such tolerance can be modeled as a (m, k) constraint which requires at least m correct runs out of any k consecutive runs to be correct. In this paper, we discuss how a given (m, k) constraint can be satisfied by adopting patterns of task instances with individual error detection and compensation capabilities. We introduce static strategies and provide a formal feasibility analysis for validation. Furthermore, we develop an adaptive scheme that extends our initial approach with online awareness that increases efficiency while preserving analysis results. The effectiveness of our method is shown in a real-world case study as well as for synthesized task sets.},
}
To avoid catastrophic events like unrecoverable system failures on mobile and embedded systems caused by soft-errors, software-
based error detection and compensation techniques have been proposed. Methods like error-correction codes or redundant execution can offer high flexibility and allow for application-specific fault-tolerance selection without the needs of special hardware supports. However, such software-based approaches may lead to system overload due to the execution time overhead. An adaptive deployment of such techniques to meet both application requirements and system constraints is desired. From our case study, we observe that a control task can tolerate limited errors with acceptable performance loss. Such tolerance can be modeled as a (m, k) constraint which requires at least m correct runs out of any k consecutive runs to be correct. In this paper, we discuss how a given (m, k) constraint can be satisfied by adopting patterns of task instances with individual error detection and compensation capabilities. We introduce static strategies and provide a formal feasibility analysis for validation. Furthermore, we develop an adaptive scheme that extends our initial approach with online awareness that increases efficiency while preserving analysis results. The effectiveness of our method is shown in a real-world case study as well as for synthesized task sets.
|
| Jakob Richter, Helena Kotthaus, Bernd Bischl, Peter Marwedel, Jörg Rahnenführer and Michel Lang.
Faster Model-Based Optimization through Resource-Aware Scheduling Strategies.
In Proceedings of the 10th International Conference: Learning and Intelligent Optimization (LION 10). vol. 10079 of Lecture Notes in Computer Science., pages 267--273
2016
[BibTeX][PDF][Link][Abstract]@inproceedings { kotthaus/2016a,
author = {Richter, Jakob and Kotthaus, Helena and Bischl, Bernd and Marwedel, Peter and Rahnenf\"uhrer, J\"org and Lang, Michel},
title = {Faster Model-Based Optimization through Resource-Aware Scheduling Strategies},
booktitle = {Proceedings of the 10th International Conference: Learning and Intelligent Optimization (LION 10).},
year = {2016},
volume = {vol. 10079 of Lecture Notes in Computer Science.},
pages = {267--273},
publisher = {Springer},
url = {http://link.springer.com/chapter/10.1007/978-3-319-50349-3_22},
file = {http://ls12-www.cs.tu-dortmund.de/daes/media/documents/publications/downloads/2016_kotthaus_lion10.pdf },
confidential = {n},
abstract = {We present a Resource-Aware Model-Based Optimization framework RAMBO that leads to efficient utilization of parallel computer architectures through resource-aware scheduling strategies. Conventional MBO fits a regression model on the set of already evaluated configurations and their observed performances to guide the search. Due to its inherent sequential nature, an efficient parallel variant can not directly be derived, as only the most promising configuration w.r.t. an infill criterion is evaluated in each iteration. This issue has been addressed by generalized infill criteria in order to propose multiple points simultaneously for parallel execution in each sequential step. However, these extensions in general neglect systematic runtime differences in the configuration space which often leads to underutilized systems. We estimate runtimes using an additional surrogate model to improve the scheduling and demonstrate that our framework approach already yields improved resource utilization on two exemplary classification tasks.},
}
We present a Resource-Aware Model-Based Optimization framework RAMBO that leads to efficient utilization of parallel computer architectures through resource-aware scheduling strategies. Conventional MBO fits a regression model on the set of already evaluated configurations and their observed performances to guide the search. Due to its inherent sequential nature, an efficient parallel variant can not directly be derived, as only the most promising configuration w.r.t. an infill criterion is evaluated in each iteration. This issue has been addressed by generalized infill criteria in order to propose multiple points simultaneously for parallel execution in each sequential step. However, these extensions in general neglect systematic runtime differences in the configuration space which often leads to underutilized systems. We estimate runtimes using an additional surrogate model to improve the scheduling and demonstrate that our framework approach already yields improved resource utilization on two exemplary classification tasks.
|
| Helena Kotthaus, Jakob Richter, Andreas Lang, Michel Lang and Peter Marwedel.
Resource-Aware Scheduling Strategies for Parallel Machine Learning R Programs through RAMBO.
In Abstract Booklet of the International R User Conference (UseR!) 195
USA, Stanford, June 2016
[BibTeX][Link][Abstract]@inproceedings { kotthaus:2016b,
author = {Kotthaus, Helena and Richter, Jakob and Lang, Andreas and Lang, Michel and Marwedel, Peter},
title = {Resource-Aware Scheduling Strategies for Parallel Machine Learning R Programs through RAMBO},
booktitle = {Abstract Booklet of the International R User Conference (UseR!)},
year = {2016},
number = {195},
address = {USA, Stanford},
month = {June},
url = {http://user2016.org/files/abs-book.pdf},
confidential = {n},
abstract = {We present resource-aware scheduling strategies for parallel R programs leading to efficient utilization of parallel computer architectures by estimating resource demands. We concentrate on applications that consist of independent tasks.
The R programming language is increasingly used to process large data sets in parallel, which requires a high amount of resources. One important application is parameter tuning of machine learning algorithms where evaluations need to be executed in parallel to reduce runtime. Here, resource demands of tasks heavily vary depending on the algorithm configuration. Running such an application in a naive parallel way leads to inefficient resource utilization and thus to long runtimes. Therefore, the R package “parallel” offers a scheduling strategy, called “load balancing”.
It dynamically allocates tasks to worker processes. This option is recommended when tasks have widely different computation times or if computer architectures are heterogeneous. We analyzed memory and CPU utilization of parallel applications with our TraceR profiling tool and found that the load balancing mechanism is not sufficient for parallel tasks with high variance in resource demands. A scheduling strategy needs to know resource demands of a task before execution to efficiently map applications to available resources.
Therefore, we build a regression model to estimate resource demands based on previous evaluated tasks. Resource estimates like runtime are then used to guide our scheduling strategies. Those strategies are integrated in our RAMBO (Resource-Aware Model-Based Optimization) Framework. Compared to standard mechanisms of the parallel package our approach yields improved resource utilization.},
}
We present resource-aware scheduling strategies for parallel R programs leading to efficient utilization of parallel computer architectures by estimating resource demands. We concentrate on applications that consist of independent tasks.
The R programming language is increasingly used to process large data sets in parallel, which requires a high amount of resources. One important application is parameter tuning of machine learning algorithms where evaluations need to be executed in parallel to reduce runtime. Here, resource demands of tasks heavily vary depending on the algorithm configuration. Running such an application in a naive parallel way leads to inefficient resource utilization and thus to long runtimes. Therefore, the R package “parallel” offers a scheduling strategy, called “load balancing”.
It dynamically allocates tasks to worker processes. This option is recommended when tasks have widely different computation times or if computer architectures are heterogeneous. We analyzed memory and CPU utilization of parallel applications with our TraceR profiling tool and found that the load balancing mechanism is not sufficient for parallel tasks with high variance in resource demands. A scheduling strategy needs to know resource demands of a task before execution to efficiently map applications to available resources.
Therefore, we build a regression model to estimate resource demands based on previous evaluated tasks. Resource estimates like runtime are then used to guide our scheduling strategies. Those strategies are integrated in our RAMBO (Resource-Aware Model-Based Optimization) Framework. Compared to standard mechanisms of the parallel package our approach yields improved resource utilization.
|
| Ingo Korb, Helena Kotthaus and Peter Marwedel.
mmapcopy: Efficient Memory Footprint Reduction using Application Knowledge.
In Proceedings of the 31st Annual ACM Symposium on Applied Computing
Pisa, Italy, 2016
[BibTeX][PDF][Abstract]@inproceedings { korb:2016:sac,
author = {Korb, Ingo and Kotthaus, Helena and Marwedel, Peter},
title = {mmapcopy: Efficient Memory Footprint Reduction using Application Knowledge},
booktitle = {Proceedings of the 31st Annual ACM Symposium on Applied Computing },
year = {2016},
address = {Pisa, Italy},
file = {http://ls12-www.cs.tu-dortmund.de/daes/media/documents/publications/downloads/2016-korb-mmapcopy.pdf},
confidential = {n},
abstract = {Memory requirements can be a limiting factor for programs dealing with large data structures. Especially interpreted programming languages that are used to deal with large vectors like R suffer from memory overhead when copying such data structures. Avoiding data duplication directly in the application can reduce the memory requirements. Alternatively, generic kernel-level memory reduction functionality like deduplication and compression can lower the amount of memory required, but they need to compensate for missing application knowledge by utilizing more CPU time, leading to excessive overhead. To allow new optimizations based
on the application’s knowledge about its own memory utilization, we propose to introduce a new system call. This system call uses the existing copy-on-write functionality of the Linux kernel to avoid duplicating memory when data is copied. Our experiments using real-world benchmarks written in the R language show that our approach can yield significant improvement in CPU time compared to Kernel Samepage Merging without compromising the amount of memory saved.
},
}
Memory requirements can be a limiting factor for programs dealing with large data structures. Especially interpreted programming languages that are used to deal with large vectors like R suffer from memory overhead when copying such data structures. Avoiding data duplication directly in the application can reduce the memory requirements. Alternatively, generic kernel-level memory reduction functionality like deduplication and compression can lower the amount of memory required, but they need to compensate for missing application knowledge by utilizing more CPU time, leading to excessive overhead. To allow new optimizations based
on the application’s knowledge about its own memory utilization, we propose to introduce a new system call. This system call uses the existing copy-on-write functionality of the Linux kernel to avoid duplicating memory when data is copied. Our experiments using real-world benchmarks written in the R language show that our approach can yield significant improvement in CPU time compared to Kernel Samepage Merging without compromising the amount of memory saved.
|
DAES