![[ACM logo]](pics/acmlogo.gif)
Operating Systems Principles
December 12-15, 1999
Kiawah Island Resort, near
Charleston, SC, USA
Sponsored by
Association for Computing Machinery
Special Interest Group on Operating
Systems
![[Kiawah logo]](pics/palmetto-small.gif)
|
17th ACM Symposium on
Operating Systems Principles December 12-15, 1999
Sponsored by
|
|
Emmanuel CECCHET
Ph. D. Student INPG/INRIA(*) - SIRAC Laboratory
655, Avenue de l'Europe
38330 Montbonnot Saint-Martin
FRANCE
Emmanuel.Cecchet@inrialpes.fr
(*) INPG : Institut National Polytechnique de Grenoble
INRIA : Institut National de Recherche en Informatique et Automatique
This research focuses on using new high bandwidth and low latency
memory-mapped networks to build high performance cluster computing
servers. The SIRAC laboratory has set up a cluster of PC
interconnected by a Scalable Coherent Interface (SCI) network. We have
developed a distributed shared memory (DSM) called SciFS, tightly
integrated with the operating system, that tries to benefit from the
high performances and the remote addressing capabilities of SCI. We
briefly present our prototype then we discuss the SCI scalability
issues by commenting on performance numbers for different network
architectures. Finally we introduce the new hardware and software
problems for large SCI clusters.
Han Kiliccote Institute for Complex Engineered Systems Carnegie Mellon University 5000 Forbes Avenue Carnegie Mellon University Pittsburgh, PA 15213 kiliccote@cmu.eduPASIS is an innovative framework for demonstrating perpetually available information systems that guarantee the survivability of information under malicious attacks or system component failures. PASIS is based on a novel architecture which breaks all information into "chunks" and distributes these "information chunks" in novel ways by using information replication and dispersal methods. This enables PASIS to not have any single point of failure (i.e., it is not possible to destroy the information in PASIS or to degrade the performance, by eliminating or capturing few selected components or information chunks within the system) and thereby achieve a very high degree of security and resiliency against failures and attacks.
Yasushi Negishi IBM Research, Tokyo Laboratory 1623-14, Shimotsuruma, Yamato-shi, Kanagawa-ken, 242-8502, Japan negishi@trl.ibm.com (or negishi@jp.ibm.com)Tuplink is a communication system for supporting requirements of communication systems for PDAs and micro-devices, those are to support operations while the link is disconnected, to work with limited memory and to use low quality links causing accidental packet loss and sudden disconnection. Tuplink is a Linda-based system, and supports one-to-one communication. Two pools are placed at both ends of communication path. Linda operations are used for handling each of local pool, and "sync" operation is introduced to synchronize the contents of the two pools. We also explain Tuplink protocol and a sample application using Tuplink, called "microNotes". MicroNotes is a commercial software supporting subset of Lotus Notes functions. Demonstration of microNotes is also shown.
Anurag Acharya, Maximilian Ibel, Matthias Koelsch, and Michael Schmitt Department of Computer Science UC Santa Barbara Santa Barbara, CA 93106 ibel@cs.ucsb.eduThe demand for net-based services continues to expand. Writing such services remains hard as each such service usually needs to provide fault tolerance, service discovery, and scalability. Typically, this functionality is implemented from scratch for every service. There is, however, significant commonality between the requirements for large class of services. The goal of the RENS project is to provide this common functionality in a runtime environment, thus freeing service designers to focus on service-specific code. RENS services consist of distributed components connected by named communication channels. The communication structure of RENS services is specified explicitly, and components are programmed in an event-driven manner, having only knowledge of local state. The RENS runtime environment leverages its knowledge of the communication structure to provide automatic instantiation of distributed services, detection of and transparent recovery from common faults, and scaling on demand. A RENS prototype is currently operational and has been used to develop a suite of distributed services including location dependent information servers, customcast, selectcast, anycast, chat servers, a hierarchical streaming audio network and caching proxies for intermittently connected devices.
Ian McDonald Department of Computing Science University of Glasgow 17 Lilybank Gardens Glasgow G12 8QQ ian@dcs.gla.ac.ukOngoing work is presented that takes advantage of several key trends in computing science to provide a novel, flexible, extensible, distributed, user-level virtual memory management hierarchy offering quality of service (QoS) to applications. The new hierarchy incorporates compressed caching, disk pre-fetching and remote paging to provide a flexible environment that allows application developers to finely tune their run-time environment. An initial implementation of this new hierarchy has been built, with future work focusing on providing a high level QoS interface.
Paal Halvorsen, Thomas Plagemann, Vera Goebel UniK - Center for Technology at Kjeller, University of Oslo, Postbox 70, N-2027 Kjeller Norway paalh@unik.noTraditional operating systems do not provide adequate support for a multimedia-on-demand server supporting video and audio playout to a large amount of users. The objective of the INSTANCE project is to identify and eliminate possible bottlenecks of traditional operating systems by improving the design of traditional server based systems. So far, we have come up with three possible orthogonal system improvements: network level framing, integrated error management and a zero-copy-one-copy memory architecture. In this poster, we present the basic ideas of our system improvements.
Jon Howell and David Kotz
Dartmouth College
6211 Sudikoff Lab
Hanover, NH 03755
{jonh,dfk}@cs.dartmouth.edu
Restricted delegation allows transitive sharing, a requirement for
sharing that spans administrative boundaries. Each principal in a chain
of delegation can restrict what is shared with the next, so that each
principal can act both as a user (recipient) and administrator (giver)
of any resource. A logic and formal semantics helps us intuitively
grasp the consequences of a security model. They justify the decision
procedure that ends up in the implementation, and they can confirm or
reject the soundness of proposed extensions. Our logic and semantics
adds restricted delegation to the calculus for access control while
preserving first-class conjunct and quoting principals.
Rolf Neugebauer Department of Computing Science University of Glasgow Glasgow, G12 8QQ, Scotland, U.K neugebar@dcs.gla.ac.ukThis poster presents a Unix-like run-time environment for the Nemesis operating system. Its aim is to provide Nemesis' advanced resource management mechanisms to existing Unix applications with no or few source code changes. In this poster we describe the overall design goals and then focus on some interesting aspects of their implementation.
Prashant Pradhan, Tzi-cker Chiueh Experimental Computer Systems Lab Dept. of Computer Science SUNY at Stony Brook Stony Brook, NY-11794-4400 prashant@cs.sunysb.eduIncreasing demand for network bandwidth and growing need for richer functionality within the network, have given rise to conflicting requirements for network router design. Edge routers, in particular, are faced with the challenge of maintaining high-speed forwarding performance, as well as providing versatile functionality. Suez is a scalable cluster-based architecture designed to provide high raw packet forwarding performance and flexible protocol processing extensibility at edge routers. The implementation of the Suez architecture consists of two major components:
Fay Chang and Garth Gibson School of Computer Science Carnegie Mellon University 5000 Forbes Avenue Pittsburgh, PA 15213 fwc@cs.cmu.eduAggressive prefetching is an effective technique for improving the performance of disk-bound applications. However, it requires timely and accurate knowledge of the future data needs of applications. We propose that applications can discover their own future data needs by exploiting unused processing cycles during I/O stalls to speculatively pre-execute their code. We find that a naive design for transforming application binaries to perform speculative execution obtains significant performance benefits in many cases, but falls short of the benefits of manually modifying applications to provide prefetching information. Early experiments show that we can enhance the effectiveness of our approach by combining static and dynamic techniques to make speculation self-aware.
Khalil Amiri, David Petrou, Greg Ganger, Garth Gibson Carnegie Mellon University 5000 Forbes Avenue Pittsburgh, PA 15213 khalil.amiri@cs.cmu.eduTraditional distributed filesystems and applications statically partition their functions between client and server machines. Dynamic changes in workload and system characteristics often make this design-time function placement suboptimal. Our prototype system, ABACUS, improves performance and manageability by automatically migrating application (e.g., search) and filesystem function (e.g., caching, RAID, directory service) among cluster nodes at run-time. Our initial experiments show that dynamic placement results in significant benefits in response time, often outperforming the best possible static placement and requiring only black-box monitoring of functional modules (objects) and of the resources available throughout the cluster.
Steven D. Gribble, Eric A. Brewer, David Culler, and Joseph M. Hellerstein UC Berkeley CS Division 445 Soda Hall #1776 UC Berkeley Campus, Berkeley, CA 94720-1776 gribble@cs.berkeley.eduIn this work, we propose a new abstraction for persistent data management on clusters of workstations: "distributed data structures", or DDS. A DDS is a scalable, self-managing repository for persistent, consistent, highly-available data that exports a data-structure interface (such as a hash table, log, or tree). A DDS decouples consistency and persistence from service logic, simplifying service construction. The poster discusses the design and implementation of a hash-table DDS, including the use of asynchronous state-machine driven I/O, and describes two example services and future directions.
Surendar Chandra, Carla Schlatter Ellis, Amin Vahdat Department of Computer Science, Duke University, PO Box 90129 Department of Computer Science, Duke University, Durham, NC 27708-0129 surendar@cs.duke.eduThis poster addresses ongoing work on a web server that manages its consumed resources and provides differentiated Quality of Service (QoS) using a quality-aware image transcoding. The key insight behind our work is that allocation of critical resources can be done dynamically at the application level in response to client access patterns using informed transcoding techniques with well understood tradeoff characteristics. These techniques enable us to control and allocate the system resources to the various client classes (in a loaded web server). We use informed transcoding techniques to provide different variations of the same multimedia object. Preliminary results are encouraging; we can not only manage the total consumed resources but also provide differentiated QoS. Our system also avoids unnecessary transcoding operations and unproductive decreases in content quality.
Donald Miller and Alan Skousen Computer Science and Engineering Department Arizona State University Tempe, AZ 85287-5406 donald.miller@asu.eduThe Sombrero operating system project is an experiment to investigate the potential of a distributed large Single Address Space Operating System (SASOS) as a computing environment. We are currently prototyping Sombrero on Alpha 21164 based NT systems. The prototype gives NT the ability to manage 8-terabyte data sets across a network in a single distributed address space. This distributed address space supports a mechanism that allows multiple network users to have access to and a consistent view of the same addresses while at the same time being protected from one another's activities. We also propose changes to current stock processor architecture to facilitate the single large distributed address space design. Sombrero provides a flat networkwide peer level architecture for user and system programs. Sombrero makes use of its hardware support for object grained protection, separate thread related protection domains and implicit protection domain crossing to provide support for modern distributed computing that is not available from SASOSs built on stock processors.
Erez Zadok and Jason Nieh
Computer Science Department
Columbia University, Mail Code 0401
1214 Amsterdam Avenue
New York, NY 10027
{ezk,nieh}@cs.columbia.edu
We propose a new language, FiST, to describe stackable file systems. FiST
uses operations common to file system interfaces. From a single
description, FiST's compiler produces file system modules for multiple
platforms. The generated code handles many kernel details, freeing
developers to concentrate on the main issues of their file systems. We
built several file systems using FiST on Solaris, FreeBSD, and Linux. Our
experiences with these examples shows the following benefits of FiST:
development time is reduced more than ten times; average code size is
reduced 5--36 times; minimal performance overhead of stacking is less than
4%.
Eric Jul, Povl Koch, Jørgen S. Hansen, Michael Svendsen, Kim Henriksen, Kenn Nielsen, and Mads Dydensborg Department of Computer Science University of Copenhagen Universitetsparken 1 DK-2100 Copenhagen DENMARK zorgylp@diku.dkThe goal of this project is to combine the theory of DSM systems with database technology to build a distributed database solution that can be used by search engines and E-commerce systems where access latency and availability are important and where updates to data should be reflected in the system as soon as possible. This project focus is on centralised cluster-based Internet servers with very large data sets -- normally Tera bytes of data. The system uses the SCI technology as communication backbone, a technology that allows nodes to map memory regions of other nodes into their local address space. The SCI network handles consistency of the memory regions on the nodes.
Joon Suan Ong, Yvonne Coady and Michael J. Feeley University of British Columbia 2366 Main Mall, University of British Columbia, Vancouver, B.C. V6T 1Z4, Canada jsong@cs.ubc.caCommunication systems use zero-copy protocols and user-level networking to improve performance. They require explicit page pinning and address translation for the NI to access user-pages. By exploiting spatial locality, statically pre-pinning and lazily unpinning pages can amortize the cost of dynamic pinning. However, this leads to poor memory utilization and expensive system calls. We propose modifying the VM system to provide the NI with translation information. The NI controls pinning for DMA operations. The VM system can still exercise its MM policies over those pages.
The selection process will aim to pick topics that look like they will be of interest to the attendees, and generate lively discussions.
A poster is an ideal way to present your ongoing work in a more structured manner than a work-in-progress talk, yet one that makes for easy interaction and discussion. You might consider providing an overview of some of your recent work, or describing one aspect in more detail, or use it to get feedback on work you are thinking about starting.
To obtain a poster-session slot, please:
Your poster will be listed on the conference web site, and you'll be given the chance to provide an extended abstract (or pointer to one) for the web site, too, if you wish.
Last modified: Mon Dec 20 11:51:09 MET 1999