The Impact of File Systems on MPI-IO Scalability

Latham, Robert; Ross, Robert; Thakur, Rajeev

doi:10.1007/978-3-540-30218-6_18

Cited by 29 publications

(20 citation statements)

References 7 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…As ever larger machines with more disks are being deployed [23], a single metadata server is no longer sufficient to handle the workload. Some of the metadata issues can be ameliorated by using collective interfaces, such as MPI-I/O [21], at the clients [9]. These techniques limit accesses to one client, with results broadcast outside of the file system to the other cooperating clients.…”

Section: Introductionmentioning

confidence: 99%

File Creation Strategies in a Distributed Metadata File System

Devulapalli

Ohio

2007

2007 IEEE International Parallel and Distributed Processing Symposium

View full text Add to dashboard Cite

As computing breaches petascale limits both in processor performance and storage capacity, the only way that current and future gains in performance can be achieved is by increasing the parallelism of the system. Gains in storage performance remain low due to the use of traditional distributed file systems such as NFS, where although multiple clients can access files at the same time, only one node can serve files to the clients. New file systems that distribute load across multiple data servers are being developed; however, most implementations concentrate all the metadata load at a single server still. Distributing metadata load is important to accommodate growing numbers of more powerful clients.Scaling metadata performance is more complex than scaling raw I/O performance, and with distributed metadata the complexity increases further. In this paper we present strategies for file creation in distributed metadata file systems. Using the PVFS distributed file system as our testbed, we present designs that are able to reduce the message complexity of the create operation and increase performance. Compared to the basecase create protocol implemented in PVFS, our design delivers near constant operation latency as the system scales, does not degenerate under high contention situations, and increases throughput linearly as the number of metadata servers increase. The design schemes are applicable to any distributed file system implementation.

show abstract

Section: Introductionmentioning

confidence: 99%

File Creation Strategies in a Distributed Metadata File System

Devulapalli

Ohio

2007

2007 IEEE International Parallel and Distributed Processing Symposium

View full text Add to dashboard Cite

show abstract

“…The term nonconflicting write was introduced by Rob Ross in the specification for PVFS2 [LRT05]. While write operations to disjunct file regions behave similarly to POSIX, the results of simultaneous writes to overlapping file regions are undefined.…”

Section: Nonconflicting Write Semanticsmentioning

confidence: 99%

Analyzing Metadata Performance in Distributed File Systems

Biardzki

Ludwig

2009

Lecture Notes in Computer Science

View full text Add to dashboard Cite

Distributed file systems are important building blocks in modern computing environments.The challenge of increasing I/O bandwidth to files has been largely resolved by the use of parallel file systems and sufficient hardware. However, determining the best means by which to manage large amounts of metadata, which contains information about files and directories stored in a distributed file system, has proved a more difficult challenge.The objective of this thesis is to analyze the role of metadata and present past and current implementations and access semantics. Understanding the development of the current file system interfaces and functionality is a key to understanding their performance limitations.Based on this analysis, a distributed metadata benchmark termed DMetabench is presented.DMetabench significantly improves on existing benchmarks and allows stress on metadata operations in a distributed file system in a parallelized manner. Both intra-node and inter-node parallelity, current trends in computer architecture, can be explicitly tested with DMetabench. This is due to the fact that a distributed file system can have different semantics inside a client node rather than semantics between multiple nodes.As measurements in larger distributed environments may exhibit performance artifacts difficult to explain by reference to average numbers, DMetabench uses a time-logging technique to record time-related changes in the performance of metadata operations and also protocols additional details of the runtime environment for post-benchmark analysis.Using the large production file systems at the Leibniz Supercomputing Center (LRZ) in Munich, the functionality of DMetabench is evaluated by means of measurements on different distributed file systems. The results not only demonstrate the effectiveness of the methods proposed but also provide unique insight into the current state of metadata performance in modern file systems.The thesis concludes with a discussion of the results and the identification of areas for future research. ii

show abstract

“…The performance of these management operations is important to the parallel applications. To evaluate the strength of different interconnects to the performance of these management operations, we have performed the following experiments using a microbenchmark program [16], op test, in the PVFS2 [2] distribution.…”

Section: Parallel Management Operationsmentioning

confidence: 99%