Proceedings of the 22nd European MPI Users' Group Meeting 2015
DOI: 10.1145/2802658.2802662
|View full text |Cite
|
Sign up to set email alerts
|

On the Impact of Synchronizing Clocks and Processes on Benchmarking MPI Collectives

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
2
1
1
1

Citation Types

0
12
0
1

Year Published

2017
2017
2023
2023

Publication Types

Select...
4
3
1

Relationship

0
8

Authors

Journals

citations
Cited by 11 publications
(13 citation statements)
references
References 11 publications
0
12
0
1
Order By: Relevance
“…Our objective is to utilize the SKaMPI framework for benchmarking OpenSHMEM communication. However, considering studies of potential clock drift [9], we know that both barrier and window-based process synchronization suffer from drift and the processes might lose their synchronization as the measurement progresses. SKaMPI's window-based clock synchronization can measure operations very accurately, but the logical global clocks drift quickly, so only a small number of MPI operations can be measured precisely.…”
Section: Measuring Openshmem Communication Routinesmentioning
confidence: 99%
See 1 more Smart Citation
“…Our objective is to utilize the SKaMPI framework for benchmarking OpenSHMEM communication. However, considering studies of potential clock drift [9], we know that both barrier and window-based process synchronization suffer from drift and the processes might lose their synchronization as the measurement progresses. SKaMPI's window-based clock synchronization can measure operations very accurately, but the logical global clocks drift quickly, so only a small number of MPI operations can be measured precisely.…”
Section: Measuring Openshmem Communication Routinesmentioning
confidence: 99%
“…SKaMPI's window-based clock synchronization can measure operations very accurately, but the logical global clocks drift quickly, so only a small number of MPI operations can be measured precisely. The hierarchical algorithm presented in [9] has a smaller clock drift, but the processes still skew during the measurement. As a consequence, we cannot rely on this synchronization only to perform our measurements, and whenever possible, we need to design measurement strategies that rely on more precise measurements than just a synchronization followed by a call to the measured routine.…”
Section: Measuring Openshmem Communication Routinesmentioning
confidence: 99%
“…The synchronization of the clocks does not produce a significant computation nor communication overhead since the clocks are synchronized once when MPI_Init is called, for the rest of the application's execution time. The effects of time drift among synchronized clocks and techniques used to keep clocks in sync are studied in [27] and [28] which are taken into account by the MPI implementation use for our proposed strategy. The timestamps for the toc are obtained with an MPI call to MPI_Wtime by each p i , which returns a number of seconds (not clock ticks) that represent elapsed wall-clock time, with a precision of 1 millionth of a second 3 .…”
Section: Toc-w: Timestamp-based Ordering For Concurrent Writingmentioning
confidence: 99%
“…idle time), as communication time. We should note that although the MPI_Barrier is commonly utilized for process synchronization [Hunold and Carpen-Amarie, 2015], its synchronization quality depends on the system and alternative synchronization methods can provide better measurement quality. Each process measures its local communication time for each iteration and reports the 3rd quantile of the times measured across all iterations.…”
Section: Benchmarkingmentioning
confidence: 99%
“…Στην πράξη, η χρήση πληροφορίας από το υποκείμενο σύστημα βελτιώνει την ακρίβεια των αναλυτικών μοντέλων και καθιστά τα ημι-εμπειρικά μοντέλα κατάλληλα για πιο περίπλοκες αποφάσεις σχετικά με το συγκεκριμένο σύστημα. Τα εμπειρικά μοντέλα [Hunold and Carpen-Amarie, 2015], η ποιότητα του συγχρονισμού που προσφέρει εξαρτάται από το σύστημα. Εναλλακτικές προσεγγίσεις για το συγχρονισμό μπορούν να αποδώσουν καλύτερη ποιότητα μετρήσεων.…”
Section: Appendix Bunclassified