Multithreaded Two-Phase I/O: Improving Collective MPI-IO Performance on a Lustre File System

Tsujita, Yuichi; Yoshinaga, Kazumi; Hori, Atsushi; Sato, Mikiko; Namiki, Mitaro; Ishikawa, Yutaka

doi:10.1109/pdp.2014.46

Cited by 15 publications

(13 citation statements)

References 8 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…ROMIO, a popular MPI I/O implementation, introduces TwoPhase I/O -an algorithm that attempts to merge noncontiguous requests into larger and more contiguous [6]. Tsujita, Y. et al obtained remarkable improvements by extending Two-Phase I/O even further, by using multiple threads [7]. Other researchers are focused on new MPI I/O implementations [8] or improvements in PFS [9] [10].…”

Section: Related Workmentioning

confidence: 99%

A Parallel MPI I/O Solution Supported by Byte-addressable Non-volatile RAM Distributed Cache

Malinowski

Czarnul

Dorozynski³

et al. 2016

Annals of Computer Science and Information Systems

View full text Add to dashboard Cite

Abstract-While many scientific, large-scale applications are data-intensive, fast and efficient I/O operations have become of key importance for HPC environments. We propose an MPI I/O extension based on in-system distributed cache with data located in Non-volatile Random Access Memory (NVRAM) available in each cluster node. The presented architecture makes effective use of NVRAM properties such as persistence and byte-level access behind the MPI I/O API. Another advantage of the proposed solution is making development of a parallel application easy and efficient as a programmer just needs to use the well known MPI I/O data model and API while efficient file access is automatically provided without a need for application level optimizations like avoiding frequent operations on a small data.Results of experiments obtained with three different applications suggest, that the extension significantly reduces file access time, especially for small I/O operations. By locating cache facilities on computing nodes, the extension decreases load of file system servers and makes I/O scalable.

show abstract

Section: Related Workmentioning

confidence: 99%

A Parallel MPI I/O Solution Supported by Byte-addressable Non-volatile RAM Distributed Cache

Malinowski

Czarnul

Dorozynski³

et al. 2016

Annals of Computer Science and Information Systems

View full text Add to dashboard Cite

show abstract

“…This approach splits the total contiguous data access regions on a file evenly between the processes. Tsujita et al [2] have optimized the TP-IO scheme by employing separate threads for data exchange and I/O operations, thereby overlapping these two operations. However Dickens and Thakur [3] observe that overlapping of I/O and computation by using background threads for collective I/O and a foreground thread for computation is often a worst implementation option.…”

Section: Related Workmentioning

confidence: 99%

Multithreaded implementation of data intensive applications with overlapped I/O and computation

Raju¹,

Chiplunkar²,

Karanth³

2015

2015 National Conference on Parallel Computing Technologies (PARCOMPTECH)

View full text Add to dashboard Cite

The performance of parallel systems is limited by the I/O performance rather than the CPU performance. Scientific computing and database applications demand for large volume of I/O especially the disk I/O. The widening gap between CPU and disk speeds limits the speedup achieved by parallelizing such I/O intensive applications. Hence reducing I/O overhead in parallel systems deserves special attention.In this paper we introduce a technique that overlaps the disk I/O and computation for an OpenMP based parallel processing environment. The technique uses multiple threads that perform computation in parallel and a single thread for I/O. The overlapping is enabled by using pre-fetch buffers for every computational thread. The above method is tested for a weather data analysis application and the experimental results show that this approach can achieve a speed up of 1.75 over a single threaded implementation. Our system achieves a speed up as high as 8.71 over the Hadoop implementation of the same application.Keywords-Parallel systems, overlapping of disk I/O and computation, pre-fetch buffers, OpenMP, weather data analysis, Hadoop.

show abstract

“…With an increase in the number of I/O request slots in queues, we may have a higher overlap effect, and so we may need larger memory spaces proportional to the number of slots. Therefore, we need to find an appropriate number of I/O request slots in terms of both throughput gains and effective We already implemented such a multithreaded scheme on a Linux PC cluster system [7]. Unlike Fig.…”

Section: Multithreaded Approach In Two-phase I/omentioning

confidence: 99%

“…We already implemented a multithreaded scheme [7] for collective write operations using Pthreads [8]. An I/O thread is invoked by every MPI process to realize overlapped operations, and the I/O thread is responsible for file I/O while the main thread is responsible for access pattern checks and data exchanges.…”

Section: Introductionmentioning

confidence: 99%

Affinity-aware optimization of multithreaded two-phase I/O for high throughput collective I/O

Tsujita¹,

Hori

Ishikawa

2014

2014 International Conference on High Performance Computing &Amp; Simulation (HPCS)

Self Cite

View full text Add to dashboard Cite

Collective MPI-IO for non-contiguous accesses has been playing a big role in not only direct MPI-IO API calls but also scientific applications using parallel I/O libraries such as HDF5, which utilizes MPI-IO APIs underneath its parallel I/O APIs. We have been focusing on performance improvements in such a collective MPI-IO by using a representative MPI-IO library named ROMIO. Inside ROMIO, an optimization scheme named two-phase I/O achieves higher performance even if we have non-contiguous accesses. We have developed multithreaded ROMIO using Pthreads for further performance improvement. In this paper, we present a better performance optimization in collective write operations by using a newly implemented functionality to manage CPU core bindings for invoked I/O threads in addition to a multiple I/O request queueing scheme. We achieved performance gains up to 29% with the CPU core bindings compared to I/O throughput without CPU core bindings. Furthermore, we noted that a multiple number of I/O request slots in queues mitigated the internal unbalanced data-exchange phase times among MPI processes.

show abstract

Multithreaded Two-Phase I/O: Improving Collective MPI-IO Performance on a Lustre File System

Cited by 15 publications

References 8 publications

A Parallel MPI I/O Solution Supported by Byte-addressable Non-volatile RAM Distributed Cache

A Parallel MPI I/O Solution Supported by Byte-addressable Non-volatile RAM Distributed Cache

Multithreaded implementation of data intensive applications with overlapped I/O and computation

Affinity-aware optimization of multithreaded two-phase I/O for high throughput collective I/O

Contact Info

Product

Resources

About