Detecting Application Load Imbalance on High End Massively Parallel Systems

DeRose, Luiz; Homer, Bill; Johnson, Dean R.

doi:10.1007/978-3-540-74466-5_17

Cited by 50 publications

(35 citation statements)

References 10 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…To measure the degree of load imbalance amongst parallel threads two metrics have been suggested: the relative load imbalance [23] and the imbalance percentage (IP) [22]. The former is a ratio of the deviation of the execution time of the longest running thread from the average execution time of the threads divided by the execution time of the longest running thread; the latter is a normalized version of the former with values between 0 and 100.…”

Section: Iterative Method: Load Imbalance Minimization (Minloadimb)mentioning

confidence: 99%

“…Intuitively, it is useful to see the imbalance percentage (IP) as the average percentage of time that the parallel threads are waiting at the end of a parallel section for the slowest thread to finish [22]. Once the balancing algorithm is activated, at the end of each iteration of the parallel application, it tries to remove one way (from the cache allocation) of the fastest threads and assign it to the slowest threads.…”

Section: Iterative Method: Load Imbalance Minimization (Minloadimb)mentioning

confidence: 99%

See 1 more Smart Citation

Load balancing using dynamic cache allocation

Moretó

Cazorla

Sakellariou

et al. 2010

Proceedings of the 7th ACM International Conference on Computing Frontiers

View full text Add to dashboard Cite

Supercomputers need a huge budget to be built and maintained. To maximize the usage of their resources, application developers spend time to optimize the code of the parallel applications and minimize execution time. Despite this effort, load imbalance still arises in many optimized applications due to causes not controlled by the application developer, resulting in significant performance degradation and waste of CPU time. If the nodes of the supercomputer use chip multiprocessors, this problem may become even worse, as the interaction between different threads inside the chip may affect their performance in an unpredictable way.Although there are many techniques to address load imbalance at run-time, as it happens, these techniques may not be particularly effective when the cause of the imbalance is due to the performance sensitivity of the parallel threads when accessing a shared cache. To this end, we present a novel run-time mechanism, with minimal hardware, that automatically tries to balance parallel applications using dynamic cache allocation. The mechanism detects which applications may be sensitive to cache allocation and reduces imbalance by assigning more cache space to the slowest threads. The efficiency of our proposed mechanism is demonstrated with both synthetic workloads and a realworld parallel application. In the former case, we reduce the execution time by up to 28.9%; in the latter case, our proposal reduces the imbalance of a non-optimized version of the application to the values obtained with a hand-tuned version of the same application.

show abstract

Section: Iterative Method: Load Imbalance Minimization (Minloadimb)mentioning

confidence: 99%

Section: Iterative Method: Load Imbalance Minimization (Minloadimb)mentioning

confidence: 99%

Load balancing using dynamic cache allocation

Moretó

Cazorla

Sakellariou

et al. 2010

Proceedings of the 7th ACM International Conference on Computing Frontiers

View full text Add to dashboard Cite

show abstract

“…Also, the ratio of critical-path imbalance and the total time of an activity on the critical path provides a useful load imbalance metric. Thus, it provides similar guidance as prior profilebased load imbalance metrics (e.g., the load-imbalance percentage metrics defined in CrayPat [3]), but the critical-path imbalance indicator can often draw a more accurate picture. The critical path retains dynamic effects in program execution, such as shifting of imbalance between processes over time, which per-process profiles simply cannot capture.…”

Section: Critical-path Imbalance Indicatormentioning

confidence: 99%

“…CrayPat [3] calculates imbalance metrics from profiles that provide measures of absolute and relative load imbalance. HPC-TOOLKIT [23] attributes the costs of idleness at global synchronization points to overloaded call paths, highlighting imbalances in call-path profiles.…”

Section: Related Workmentioning

confidence: 99%

Scalable Critical-Path Based Performance Analysis

Böhme

Wolf

Supinski

et al. 2012

2012 IEEE 26th International Parallel and Distributed Processing Symposium

View full text Add to dashboard Cite

Abstract-The critical path, which describes the longest execution sequence without wait states in a parallel program, identifies the activities that determine the overall program runtime. Combining knowledge of the critical path with traditional parallel profiles, we have defined a set of compact performance indicators that help answer a variety of important performance-analysis questions, such as identifying load imbalance, quantifying the impact of imbalance on runtime, and characterizing resource consumption. By replaying event traces in parallel, we can calculate these performance indicators in a highly scalable way, making them a suitable analysis instrument for massively parallel programs with thousands of processes. Case studies with real-world parallel applications confirm that-in comparison to traditional profiles-our indicators provide enhanced insight into program behavior, especially when evaluating partitioning schemes of MPMD programs.

show abstract

“…Although previous work has focused on specific balancing algorithms [15,21,22], it is difficult to directly apply them to virtualized servers. Additionally, we need to take into consideration the impact that the virtualization layer imposes on system resources.…”

Section: Load Imbalance Metricmentioning

confidence: 99%

Quantifying load imbalance on virtualized enterprise servers

Arzuaga

Kaeli

2010

Proceedings of the First Joint WOSP/SIPEW International Conference on Performance Engineering

View full text Add to dashboard Cite

Virtualization has been shown to be an attractive path to increase overall system resource utilization. The use of live virtual machine (VM) migration has enabled more effective sharing of system resources across multiple physical servers, resulting in an increase in overall performance. Live VM migration can be used to load balance virtualized clusters. To drive live migration, we need to be able to measure the current load imbalance. Further, we also need to accurately predict the resulting load imbalance produced by any migration.In this paper we present a new metric that captures the load of the physical servers and is a function of the resident VMs. This metric will be used to measure load imbalance and construct a load-balancing VM migration framework. The algorithm for balancing the load of virtualized enterprise servers follows a greedy approach, inductively predicting which VM migration will yield the greatest improvement of the imbalance metric in a particular step. We compare our algorithm to the leading commercially available load balancing solution -VMware's Distributed Resource Scheduler (DRS). Our results show that when we are able to accurately measure system imbalance, we can also predict future system state. We find that we can outperform DRS and improve performance up to 5%. Our results show that our approach does not impose additional performance impact and is comparable to the virtual machine monitor overhead.

show abstract

Detecting Application Load Imbalance on High End Massively Parallel Systems

Cited by 50 publications

References 10 publications

Load balancing using dynamic cache allocation

Load balancing using dynamic cache allocation

Scalable Critical-Path Based Performance Analysis

Quantifying load imbalance on virtualized enterprise servers

Contact Info

Product

Resources

About