Abstract not found

In this article we present results from an experimental evaluation study of the HP Exemplar file system. Our experiments consist of simple micro-benchmarks that study the impact of various factors on the file system performance. These factors include I/O request/buffer sizes, vectored/non-vectored access patterns, read-ahead policies, multi-threaded (temporally irregular) requests, and architectural issues (cache parameters, NUMA behavior, etc.). Experimental results indicate that the Exemplar file system provides high I/O bandwidth, both for single- and multi-threaded applications. The buffer cache, with prioritized buffer management and large buffer sizes, is effective in exploiting temporal and spatial access localities. The performance of noncontiguous accesses can be improved by either using vectored I/O interfaces or tuning the read-ahead facilities. The file system performance depends on the relative locations of the computing threads and the file system, and also on various Exe...

The slab allocator [Bonwick94] has taken on a life of its own since its introduction in these pages seven

Rights to individual papers remain with the author or the author's employer. Permission is granted for noncommercial reproduction of the work for educational or research purposes. This copyright notice must be included in the reproduced paper. USENIX acknowledges all trademarks herein. For more information about the USENIX Association:

In computer systems today, speed and responsiveness is often determined by network and storage subsystem performance. Faster, more scalable networking interfaces like Fibre Channel and Gigabit Ethernet provide the scaffolding from which higher performance computer systems implementations may be constructed, but new thinking is required about how machines interact with network-enabled storage devices. In this paper we describe how we implemented journaling in the Global File System (GFS), a shared-disk, cluster file system for Linux. Our previous three papers on GFS at the Mass Storage Symposium discussed our first three GFS implementations, their performance, and the lessons learned. Our fourth paper describes, appropriately enough, the evolution of GFS version 3 to version 4, which supports journaling and recovery from client failures. In addition, GFS scalability tests extending to 8 machines accessing 8 4-disk enclosures were conducted: these tests showed good scaling. W...

To reduce performance degradation from Translation Lookaside Bu#er #TLB# misses without signi#- cant increase in TLB size, most modern processors implement TLBs that support multiple pagesizes. For example, Hewlett-Packard's PA-8000 processor allows 8 hardware pagesizes, in multiples of four, ranging from 4 Kbytes to 64 Mbytes. In implementing multiple pagesize support in HP-UX, wechose to create large pages at page-fault service time. Wehave a buddy system allocator that provides interfaces for allocating and freeing multiple pagesizes. We maintain the Virtual Memory #VM# data structures such as the pagetable entry, virtual page frame descriptor, and physical page frame descriptor based on the smallest pagesize, and represent a large pagesize as a collection of these base pagesize structures. In our implementation, VM operations on a large pagesize such as 16KB are carried out by looping over the 4KB-based constituent VM data structures. Our system o#ers signi#cant application perfor...

Parallel computing is integral to high performance computing, but it is not uniquely sufficient. With the adoption of parallel computing, some additional supporting technologies are required. Parallel I/O is one such supporting technology, providing high speed data storage in parallel computing environments. Parallel I/O systems have emerged and are beginning to see use in the main stream; however, research into optimizing these systems is still an open area. In particular, techniques for optimizing parallel I/O have focused on disk performance optimization when other resources might have equal or greater impact on overall performance. Other work has looked at adaptive techniques for optimizing in these systems, but has focused on caching and prefetching only.

Providing efficient device driver support in the Fluke operating system presents novel challenges, which stem from two conflicting factors: (i) a design and maintenance requirement to reuse unmodified legacy device drivers, and (ii) the mismatch between the Fluke kernel's internal execution environment and the execution environment expected by these legacy device drivers. This thesis presents a solution to this conflict: a framework whose design is based on running device drivers as usermode servers, which resolves the fundamental execution environment mismatch. This approach

Implicit coscheduling strategies enable parallel applications to dynamically share the machines in a Network of Workstation (NOW) with interactive, CPU and IO-bound sequential jobs. In this paper we present a simulation study that compares 12 coscheduling strategies in terms of their impact on the performance of parallel and sequential applications executed simultaneously on a NOW. Our results show that the coscheduling strategy has a strong impact on the performance of the applications (both parallel and sequential) composing the workload, and that no single strategy is able to effectively handle all workloads. In spite of that, our results can be used to identify the strategy that represents the best choice for a given application class, or the best compromise for various workloads. Moreover, we show that in many cases simple strategies outperform more complex ones.

When designing an extensible operating system, a developer must ensure that the operating system is protected from misbehaved extensions. Two kinds of protection are needed: first, extensions should not violate the operating system’s interface, and second, extensions should not be able to leave the operating system in an inconsistent state. The major research contributions of this thesis include: The design and evaluation of MiSFIT, a software fault isolation tool for the x86 architecture that ensures that extensions do not violate the operating system’s interface and incurs minimal overhead. The design and evaluation of VINO Lightweight Transactions, a low-overhead mechanism that allows the kernel to maintain its consistency in the face of ill-behaved extensions. Experiments that show the end-to-end overhead of MiSFIT and VLT protection is low, on the order of 1-2%, and the net performance gain possible from using application-specific extensions is significant, in some cases more than 20%. A cost-benefit framework for comparing extension technologies and an evaluation comparing