[Also available in BibTeX] [See also: all keywords]
These papers relate to parallel I/O, that is, I/O for parallel computers.Papers are listed in reverse-chronological order;
click an entry to pop up the abstract.
For full information and pdf, please click Details link.
Follow updates with RSS.
Using zero(), it is possible to efficiently implement applications including a variety of databases and I/O-efficient computation systems on top of the Unix file system. zero() can also be used to implement an efficient file-system-based paging mechanism. In some I/O-efficient computations, the availability of zero() effectively doubles disk capacity by allowing blocks of temporary files to be reallocated to new files as they are read.
Experiments on a Linux ext2 file system augmented by zero() demonstrate that where their functionality overlaps, zero() is more efficient than ftruncate(). Additional experiments reveal that in exchange for added effective disk capacity, I/O-efficient code pays only a small performance penalty.
In this work we examine current multiprocessor file systems, as well as how those file systems are used by scientific applications. Contrary to the expectations of the designers of current parallel file systems, the workloads on those systems are dominated by requests to read and write small pieces of data. Furthermore, rather than being accessed sequentially and contiguously, as in uniprocessor and supercomputer workloads, files in multiprocessor file systems are accessed in regular, structured, but non-contiguous patterns.
Based on our observations of multiprocessor workloads, we have designed Galley, a new parallel file system that is intended to efficiently support realistic scientific multiprocessor workloads. In this work, we introduce Galley and discuss its design and implementation. We describe Galley’s new three-dimensional file structure and discuss how that structure can be used by parallel applications to achieve higher performance. We introduce several new data-access interfaces, which allow applications to explicitly describe the regular access patterns we found to be common in parallel file system workloads. We show how these new interfaces allow parallel applications to achieve tremendous increases in I/O performance. Finally, we discuss how Galley’s new file structure and data-access interfaces can be useful in practice.
The design of a high-performance multiprocessor file system requires a comprehensive understanding of the expected workload. Unfortunately, until recently, no general workload studies of multiprocessor file systems have been conducted. The goal of the CHARISMA project was to remedy this problem by characterizing the behavior of several production workloads, on different machines, at the level of individual reads and writes. The first set of results from the CHARISMA project describe the workloads observed on an Intel iPSC/860 and a Thinking Machines CM-5. This paper is intended to compare and contrast these two workloads for an understanding of their essential similarities and differences, isolating common trends and platform-dependent variances. Using this comparison, we are able to gain more insight into the general principles that should guide multiprocessor file-system design.
We propose that the traditional functionality of parallel file systems be separated into two components: a fixed core that is standard on all platforms, encapsulating only primitive abstractions and interfaces, and a set of high-level libraries to provide a variety of abstractions and application-programmer interfaces (APIs).
We present our current and next-generation file systems as examples of this structure. Their features, such as a three-dimensional file structure, strided read and write interfaces, and I/O-node programs, re specifically designed with the flexibility and performance necessary to support a wide range of applications.
We have found that the Galley File System provides a good environment on which to build high-performance libraries, and that the mesh of Panda and Galley was a successful combination.
Recent parallel file-system usage studies show that writes to write-only files are a dominant part of the workload. Therefore, optimizing writes could have a significant impact on overall performance. In this paper, we propose ENWRICH, a compute-processor write-caching scheme for write-only files in parallel file systems. ENWRICH combines low-overhead write caching at the compute processors with high performance disk-directed I/O at the I/O processors to achieve both low latency and high bandwidth. This combination facilitates the use of the powerful disk-directed I/O technique independent of any particular choice of interface. By collecting writes over many files and applications, ENWRICH lets the I/O processors optimize disk I/O over a large pool of requests. We evaluate our design via simulated implementation and show that ENWRICH achieves high performance for various configurations and workloads.
We propose that the traditional functionality of parallel file systems be separated into two components: a fixed core that is standard on all platforms, encapsulating only primitive abstractions and interfaces, and a set of high-level libraries to provide a variety of abstractions and application-programmer interfaces (APIs). We think of this approach as the “RISC” of parallel file-system design.
We present our current and next-generation file systems as examples of this structure. Their features, such as a three-dimensional file structure, strided read and write interfaces, and I/O-node programs, are specifically designed with the flexibility and performance necessary to support a wide range of applications.
Recent parallel file-system usage studies show that writes to write-only files are a dominant part of the workload. Therefore, optimizing writes could have a significant impact on overall performance. In this paper, we propose ENWRICH, a compute-processor write-caching scheme for write-only files in parallel file systems. ENWRICH combines low-overhead write caching at the compute processors with high performance disk-directed I/O at the I/O processors to achieve both low latency and high bandwidth. This combination facilitates the use of the powerful disk-directed I/O technique independent of any particular choice of interface. By collecting writes over many files and applications, ENWRICH lets the I/O processors optimize disk I/O over a large pool of requests. We evaluate our design via simulated implementation and show that ENWRICH achieves high performance for various configurations and workloads.
The design of a high-performance parallel file system requires a comprehensive understanding of the expected workload. Unfortunately, until recently, no general workload studies of parallel file systems have been conducted. The goal of the CHARISMA project was to remedy this problem by characterizing the behavior of several production workloads, on different machines, at the level of individual reads and writes. The first set of results from the CHARISMA project describe the workloads observed on an Intel iPSC/860 and a Thinking Machines CM-5. This paper is intended to compare and contrast these two workloads for an understanding of their essential similarities and differences, isolating common trends and platform-dependent variances. Using this comparison, we are able to gain more insight into the general principles that should guide parallel file-system design.
Of course, computational processes sharing a node with a file-system service may receive less CPU time, network bandwidth, and memory bandwidth than they would on a computation-only node. In this paper we begin to examine this issue experimentally. We found that high-performance I/O does not necessarily require substantial CPU time, leaving plenty of time for application computation. There were some complex file-system requests, however, which left little CPU time available to the application. (The impact on network and memory bandwidth still needs to be determined.) For applications (or users) that cannot tolerate an occasional interruption, we recommend that they continue to use only compute nodes. For tolerant applications needing more cycles than those provided by the compute nodes, we recommend that they take full advantage of both compute and I/O nodes for computation, and that operating systems should make this possible.
Most successful systems are based on a solid understanding of the characteristics of the expected workload, but until now there have been no comprehensive workload characterizations of multiprocessor file systems. We began the CHARISMA project in an attempt to fill that gap. We instrumented the common node library on the iPSC/860 at NASA Ames to record all file-related activity over a two-week period. Our instrumentation is different from previous efforts in that it collects information about every read and write request and about the mix of jobs running in the machine (rather than from selected applications).
The trace analysis in this paper leads to many recommendations for designers of multiprocessor file systems. First, the file system should support simultaneous access to many different files by many jobs. Second, it should expect to see many small requests, predominantly sequential and regular access patterns (although of a different form than in uniprocessors), little or no concurrent file-sharing between jobs, significant byte- and block-sharing between processes within jobs, and strong interprocess locality. Third, our trace-driven simulations showed that these characteristics led to great success in caching, both at the compute nodes and at the I/O nodes. Finally, we recommend supporting strided I/O requests in the file-system interface, to reduce overhead and allow more performance optimization by the file system.
Design of such high-performance parallel file systems depends on a thorough grasp of the expected workload. So far there have been no comprehensive usage studies of multiprocessor file systems. Our CHARISMA project intends to fill this void. The first results from our study involve an iPSC/860 at NASA Ames. This paper presents results from a different platform, the CM-5 at the National Center for Supercomputing Applications. The CHARISMA studies are unique because we collect information about every individual read and write request and about the entire mix of applications running on the machines.
The results of our trace analysis lead to recommendations for parallel file system design. First, the file system should support efficient concurrent access to many files, and I/O requests from many jobs under varying load condit ions. Second, it must efficiently manage large files kept open for long periods. Third, it should expect to see small requests, predominantly sequential access patterns, application-wide synchronous access, no concurrent file-sharing between jobs, appreciable byte and block sharing between processes within jobs, and strong interprocess locality. Finally, the trace data suggest that node-level write caches and collective I/O request interfaces may be useful in certain environments.
Most successful systems are based on a solid understanding of the characteristics of the expected workload, but until now there have been no comprehensive workload characterizations of multiprocessor file systems. We began the CHARISMA project in an attempt to fill that gap. We instrumented the common node library on the iPSC/860 at NASA Ames to record all file-related activity over a two-week period. Our instrumentation is different from previous efforts in that it collects information about every read and write request and about the mix of jobs running in the machine (rather than from selected applications).
The trace analysis in this paper leads to many recommendations for designers of multiprocessor file systems. First, the file system should support simultaneous access to many different files by many jobs. Second, it should expect to see many small requests, predominantly sequential and regular access patterns (although of a different form than in uniprocessors), little or no concurrent file-sharing between jobs, significant byte- and block-sharing between processes within jobs, and strong interprocess locality. Third, our trace-driven simulations showed that these characteristics led to great success in caching, both at the compute nodes and at the I/O nodes. Finally, we recommend supporting strided I/O requests in the file-system interface, to reduce overhead and allow more performance optimization by the file system.
This dissertation studies some of the file system issues needed to get high performance from parallel disk systems, since parallel hardware alone cannot guarantee good performance. The target systems are large MIMD multiprocessors used for scientific applications, with large files spread over multiple disks attached in parallel. The focus is on automatic caching and prefetching techniques. We show that caching and prefetching can transparently provide the power of parallel disk hardware to both sequential and parallel applications using a conventional file system interface. We also propose a new file system interface (compatible with the conventional interface) that could make it easier to use parallel disks effectively.
Our methodology is a mixture of implementation and simulation, using a software testbed that we built to run on a BBN GP1000 multiprocessor. The testbed simulates the disks and fully implements the caching and prefetching policies. Using a synthetic workload as input, we use the testbed in an extensive set of experiments. The results show that prefetching and caching improved the performance of parallel file systems, often dramatically.
Experiments have been conducted with an interleaved file system testbed on the Butterfly Plus multiprocessor. Results of these experiments suggest that 1) the hit ratio, the accepted measure in traditional caching studies, may not be an adequate measure of performance when the workload consists of parallel computations and parallel file access patterns, 2) caching with prefetching can significantly improve the hit ratio and the average time to perform an I/O operation, and 3) an improvement in overall execution time has been observed in most cases. In spite of these gains, prefetching sometimes results in increased execution times (a negative result, given the optimistic nature of the study).
We explore why is it not trivial to translate savings on individual I/O requests into consistently better overall performance and identify the key problems that need to be addressed in order to improve the potential of prefetching techniques in this environment.
Experiments have been conducted with an interleaved file system testbed on the Butterfly Plus multiprocessor. Results of these experiments suggest that 1) the hit ratio, the accepted measure in traditional caching studies, may not be an adequate measure of performance when the workload consists of parallel computations and parallel file access patterns, 2) caching with prefetching can significantly improve the hit ratio and the average time to perform an I/O operation, and 3) an improvement in overall execution time has been observed in most cases. In spite of these gains, prefetching sometimes results in increased execution times (a negative result, given the optimistic nature of the study).
We explore why is it not trivial to translate savings on individual I/O requests into consistently better overall performance and identify the key problems that need to be addressed in order to improve the potential of prefetching techniques in this environment.