|
First NMADS Day |
|
Home
Steering Committee
|
March 24, 2000 [Schedule]
Invited Talk There is a debate concerning the future directions that the Internet should take. Support has been building for the Next Generation Internet (NGI) Initiative, which would greatly enhance the performance and scalability of the current Internet, while also providing improved Internet access for K-12 schools and investing in new kinds of Internet uses, such as for providing government services over the network. These include safety, revenue and life-critical systems which have traditionally been developed using special-purpose computing technology. The NGI envisions a steady transition of such systems onto a commercial off-the-shelf technology base. There really isn't any alternative; only standard technologies yield cost-effective systems. But this begs the question: What needs to be done to transform the current Internet into the NGI, if critical applications will run upon it? In particular, if the current Internet is unsafe for such applications, why should we expect the NGI to be more safe? The talk suggests that as things stand, the NGI will not be safe for critical uses, but also shows how a simple "virtual overlay network" capability could be developed to plug the major deficiency. Short Talks Using Jini for Complex Systems Integration I will describe the work that we have been doing using a Jini object model as the basis for Systems Integration, in particular where those
systems are large, distributed, complex, and have legacy components. In contrast to other distributed object systems, Jini has the unique
advantage that it is protocol-independent, meaning that any communication protocol can be encapsulated in the object's stub or
proxy that a client obtains from the Lookup Service. This is immensely important for large systems that often have to support many
co-existing protocols. Evolving the Publish-Subscribe Paradigm We discuss the work in the Gryphon project at IBM TJ Watson Research, to provide robust, scalable, internet-wide event distribution for
business and personal applications. Gryphon began as an effort to extend conventional group-based publish-subscribe to content-based
publish-subscribe. It is now being extended into a system allowing clients to subscribe to "interesting" state derived from published
events and from other state. The design and implementation of Gryphon involves a challenging synthesis of concepts drawn from both database
and distributed messaging technologies. We discuss recent results, and ongoing work. Computing Communities: Transparent Distribution Middleware The heterogeneous and dynamic nature of distributed systems makes it a challenging task to design performance-portable applications. Even
when appropriate techniques have been developed (e.g., for fault-tolerant systems), they have seen restricted deployment because
of the application development barrier: such techniques have typically required applications to adhere to a different API,
ultimately limiting acceptance. The Computing Communities project is designing transparent virtualization and adaptation techniques which
aim to address these shortcomings. These techniques lower the application development barrier by being embodied in a middleware
layer that is inserted between the application and the base operating system (Windows NT in our case) using API interception techniques. We
discuss recent results and ongoing work. Robustness Testing of CORBA ORB Implementations
Integrating Asynchronous Messaging and Distributed Object
Transactions Messaging, and distributed transactions, describe two important models for building enterprise-scale software systems. Distributed
object middleware provides services for transaction processing and for asynchronous messaging. However, the integration of distributed
transactions and asynchronous messaging in distributed object systems is mostly undefined and unsupported. In this talk, we argue for a
systematic and well-defined approach to integrating distributed transactions and messaging using middleware. We present first results
from our project of developing a messaging and transaction integration facility, and outline important research topics that need
to be addressed. Perfect Failure Detection in Timed Systems with Hardware Watchdogs I address the problem of how a backup computer can decide when it has to take over the services of a primary computer. On one hand, the
backup must not take over as long as the primary is alive because in the considered replication scheme the backup is taking over some of
the network addresses of the primary. On the other hand, when the primary is crashed, the backup has to take over within a bounded
amount of time (unless it has crashed itself). To solve this problem, I define a *timed perfect* failure detector TP that is very similar
to a perfect failure detector. The main difference is that TP has to cope with computers recovering from crash failures and TP has to
detect crashes/recoveries within a maximum known time. I propose a protocol that implements TP in timed systems with three computers
(primary, backup, and a witness) if each of the three computers has a local hardware watchdog. Exemptions from Coordination and the Origins of Primary Partition
Behavior We describe a generalization of Uniform Distributed Coordination in which not just the faulty processes are exempt from participating. We
define one exemption and show it is the origin of primary-partition behavior, seen in systems that are said to exhibit virtual synchrony.
Our characterization of the exemption allows us to weaken this exemption still further to derive problems that are inherently
solvable in systems that devolve to multiple partitions. Fast Deterministic Consensus in a Noisy Environment It is well known that the consensus problem cannot be solved deterministically in an asynchronous environment, but that randomized
solutions are possible. We propose a new model we call noisy scheduling in which an adversarial schedule is perturbed randomly,
and show that in this model randomness in the environment can substitute for randomness in the algorithm. In particular, we show
that a simplified, deterministic version of Chandra's wait-free shared-memory consensus algorithm solves consensus in time at most
logarithmic in the number of active processes. The proof of Fault Tolerant HLRC Distributed Shared Memory We describe a design for integrating fault tolerance in a distributed shared memory (DSM) system based on a Home-based Lazy Release
Consistency (HLRC) protocol. Fault tolerant HLRC uses checkpointing and volatile memory logging for rollback recovery from single node
failures. HLRC has been known as a highly efficient DSM protocol, and therefore any design for fault tolerance support should not adversely
impact its performance. Our recovery support introduces low overhead during failure-free operation by using volatile memory for logging,
and by aggressively exploiting the base DSM protocol. The main goal of this work is to investigate the feasibility and efficiency of
building a fault-tolerant DSM system that can recover from failures using only independent checkpointing and fast log-based recovery. Automatic Configuration and Run-time Adaptation of Distributed
Applications Increased platform heterogeneity and varying resource availability in distributed systems motivates the design of resource-aware
applications, which ensure a desired performance level by continuously adapting their behavior to changing resource
characteristics. In this talk, we describe an application-independent adaptation framework that simplifies the design of resource-aware
applications. This framework eliminates the need for adaptation decisions to be explicitly programmed into the application by relying
on two novel components: (1) a tunability interface, which exposes A Component Description Repository One of the most powerful factors in the development of open industrial sectors is the standardization of industrial components.
This enables industry players to unbind their operations from individual suppliers, and to take advantage of an open market of
commoditized components. Take the case of the new economy now developing on the Internet. Here, the central components are
electronic services, and several standardization processes are on their way to produce common service interfaces in a wide variety of Cluster-Based WWW Servers
Prefetching Hyperlinks We develop a new method for prefetching Web pages into the browser's cache. Clients send information about hyperlink reference patterns to
Web servers, which aggregate the reference information in near-real-time and then disperse the aggregated information to all
clients, piggybacked on GET responses. The information indicates how often hyperlink URLs embedded in pages have been previously accessed
relative to the embedding page. Based on knowledge about which hyperlinks are generally popular, clients initiate prefetching of the
hyperlinks and their embedded images according to any algorithm they prefer. Both client and server may cap the prefetching mechanism's
space overhead and waste of network resources due to speculation. The result of these differences is improved prefetching lower client
latency (by 52.3\%) and less wasted network bandwidth (24.0\%). Posters DataSpace - querying and monitoring deeply networked collections in The DataSpace is a three dimensional physical space 100 kilometers above and 10 kilometers below the surface of earth that is accessible
to the network. It is addressed geographically as opposed to the current ``logical'' addressing scheme of the Internet. With the
enormous 128 bit addressing space of IPv6, one can individually address every cubic centimeter of the physical space in DataSpace
with approximately 90 bits of area code. This would include every street, building, room, basement or even drawer of a desk. The
DataSpace would thus serve as the host for the entire part of the Transparent Network Connectivity in Dynamic Cluster Environments
Improvements in microprocessor and networking performance have made networks of workstations a very attractive platform for high-end
parallel and distributed computing. However, the effective deployment of such environments requires addressing two problems not associated
with dedicated parallel machines: heterogeneous resource capabilities and dynamic availability. Achieving good performance requires that
application components be able to migrate between cluster resources and efficiently adapt to the underlying resource capabilities. An
important component of the required support is maintaining network connectivity, which directly impacts on the transparency of migration
to the application and its performance after migration. Unfortunately, existing approaches rely on either extensive operating
system modifications or new APIs to maintain network connectivity, both of which limits their wider applicability. This poster presents
the design, implementation, and performance of a transparent network connectivity layer for dynamic cluster environments. Our design uses
the techniques of API interception and virtualization to construct a transparent layer in user space; use of the layer requires no
modification either to the application or the underlying operating system and messaging layers. Our layer enables the migration of
application components without breaking network connections, and additionally permits adaptation to the characteristics of the
underlying networking substrate. Experiments with supporting a persistent socket interface in two environments---an Ethernet LAN on
top of TCP/IP, and a Myrinet LAN on top of Fast Messages---show that our approach incurs minimal overheads and can effectively select the
best substrate for implementing application communication requirements. Software DSM on PC clusters using the Virtual Interface Architecture Our goal is to provide a shared address space over a cluster of PCs using a software DSM protocol and a low-latency high-speed
interconnect supporting the Virtual Interface Architecture (VIA). VIA standardizes the interface for user-level memory mapped communication
over high-performance System Area Networks(SAN). The focus of our project is on providing a high-performance DSM system that can be
used for scientific as well as non-scientific applications like data mining and continuous media applications. We have implemented a
Home-based Lazy Release Consistent (HLRC) DSM protocol on a cluster of eight PCs connected by a VIA-based Giganet network. Our
preliminary results indicate that software DSM on VIA achieves good performance that compares well with that achieved by similar
protocols on Myrinet-based networks. Modeling Cluster Failures
Our work investigates the distribution of failure rates for a variety of cluster components, including operating system failures, software
failures, and component failures. In particular, we focus on modeling the failure rate of the operating system. Our primary method is
empirical observation of systems running in a variety of environments, ranging from desktop PCs to clusters of Suns servers
housed in strictly controlled machine rooms. Split-OS The introduction of switched-based scalable I/O architecture for servers,
like InfiniBand, and the increasing interest for "intelligent" devices with RISC processors and memory on board are creating new opportunities
for OS designers. In particular, by redesigning the OS, we may be able to improve system reliability and increase performance through the
aggressive use of device intelligence. In Split-OS, we are redesigning the traditional OS by splitting (and
possibly replicating) OS functions and state between the host and the intelligent
devices. Our current approach is to simulate this expected server architecture using a cluster of PCs interconnected by
Virtual Interface Architecture (VIA). In this setup, each PC
emulates a particular device. We are working to define a communication protocol between devices using VIA channels in
anticipation of the InfiniBand specifications release. Co-operative File Systems for
clusters A cooperative file system is a user-level distributed file system
created ad-hoc for a distributed application by merging the local file systems of the cluster nodes. The merging
is lazily performed, temporary (application lifetime), volatile (in memory, not on disks),
local to the application , and dynamically reconfigurable. The system is called cooperative
because local file systems can "help" each other by migrating or replicating hot files among them for
better load balacing and fault tolerance. Central to the design is the concept of virtual directories provide
location independent file naming and consequently support file migration. The main application for CFS are the distributed servers like file
servers, web servers, video servers, email servers etc. Today, servers are extremely loaded and their I/O activity can be easily imbalanced.
CFS supports dynamic balacing of the disc I/O load accross the cluster using file migration and replication.
We are currently implementing a CFS prototype that uses Virtual Interface
Architecture (VIA) to interconnect 8 PCs into a cluster. Our first application
is a NFS distributed server. We will be exploring various policies for migration and studying the impact of caching policies on performance. |