This paper describes a new operating system kernel, called the x-kernel, that provides an explicit architecture for constructing and composing network protocols. Our experience implementing and evaluating several protocols in the x-kernel shows that this architecture is both general enough to accommodate a wide range of protocols, yet efficient enough to perform competitively with less structured operating systems. 1 Introduction Network software is at the heart of any distributed system. It manages the communication hardware that connects the processors in the system and it defines abstractions through which processes running on those processors exchange messages. Network software is extremely complex: it must hide the details of the underlying hardware, recover from transmission failures, ensure that messages are delivered to the application processes in the appropriate order, and manage the encoding and decoding of data. To help manage this complexity, network software is divi...

We present detailed measurements of various processing overheads of the TCP/IP and UDP/IP protocol stacks on a DECstation 5000/200 running the Ultrix 4.2a operating system. These overheads include data-touching operations, such as the checksum computation and data movemen ~ which are well known to be major time consumers. In this stud y, we also considered overheads due to non-data touching operations, such as network buffer manipulation, protocol-specific processing, operating system functions, data structure manipulations (other than network buffers), and error checking. We show that when one considers realistic message size dktributions, where the majority of messages are small, the cumulative time consumed by the nondata touching overheads represents the majority of processing time. We assert that it will be difficult to significantly reduce the cumulative processing time due to non-data touching overheads. The goal of this study is to determine the relative importance of various processing overheads in network software, in particular, the TCP/IP and UDPAP protocol stacks. In the prrsg significant focus has been placed on maximizing throughput noting that “data

Nectar is a "network backplane" for use in heterogeneous multicomputers. The initial system consists of a starshaped fiber-optic network with an aggregate bandwidth of 1.6 gigabits/second and a switching latency of 700 nanoseconds. The system can be scaled up by connecting hundreds of these networks together. The Nectar architecture provides a flexible way to handle heterogeneity and task-level parallelism. A wide variety of machines can be connected as Nectar nodes and the Nectar system software allows applications to communicate at a high level. Protocol processing is off-loaded to powerful communication processors so that nodes do not have to support a suite of network protocols. We have designed and built a prototype Nectar system that has been operational since November 1988. This paper presents the motivation and goals for Nectar and describes its hardware and software. The presentation emphasizes how the goals influenced the design decisions and led to the novel aspects of Necta...

We present an I/O architecture, called Swift, that addresses the problem of data rate mismatches between the requirements of an application, storage devices, and the interconnection medium. The goal of Swift is to support high data rates in general purpose distributed systems. Swift uses a high-speed interconnection medium to provide high data rate transfers by using multiple slower storage devices in parallel. It scales well when using multiple storage devices and interconnections, and can use any appropriate storage technology, including high-performance devices such as disk arrays. To address the problem of partial failures, Swift stores data redundantly. Using the UNIX operating system, we have constructed a simplified prototype of the Swift architecture. The prototype provides data rates that are significantly faster than access to the local SCSI disk, limited by the capacity of a single Ethernet segment, or in the case of multiple Ethernet segments by the ability of the client to drive them. We have constructed a simulation model to demonstrate how the Swift architecture can exploit advances in processor, communication and storage technology. We consider the effects of processor speed, interconnection capacity, and multiple storage agents on the utilization of the components and the data rate of the system. We show that the data rates scale well in the number of storage devices, and that by replacing the most highly stressed components by more powerful ones the data rates of the entire system increase significantly.

Over the last decade Internet has grown by orders of magnitude in size. Many of the protocols that were designed several years ago are still in use. It is not clear if the assumptions made in the design of control schemes still hold, particularly when we consider end-to-end behavior of paths in the network, today. This paper describes a simple experiment designed to capture end-to-end behavior of the Internet. The measurements indicate that the IP level service provided in the network yields high losses, duplicates and reorderings of packets. In addition, the round-trip transit delay varies significantly. These measurements indicate that the network may have several problems which still need to be analysed in order to improve the efficiency of protocols and control mechanisms that it uses.

We present the results of a study of Internet round-trip delay. The links chosen include links to frequently accessed commercial hosts as well as well-known academic and foreign hosts. Each link was studied for a 48-hour period. We attempt to answer the following questions: (1) how rapidly and in what manner does the delay change -- in this study, we focus on medium-grain (seconds/minutes) and coarse-grain time-scales (tens of minutes/hours); (2) what does the frequency distribution of delay look like and how rapidly does it change; (3) what is a good metric to characterize the delay for the purpose of adaptation. Our conclusions are: (a) there is large temporal and spatial variation in round-trip time (RTT); (b) RTT distribution is usually unimodal and asymmetric and has a long tail on the right hand side; (c) RTT observations in most time periods are tightly clustered around the mode; (d) the mode is a good characteristic value for RTT distributions; (e) RTT distributions change slow...

This paper presents detailed measurements of processing overheads for the Ultrix 4.2a implementation of TCP/IP network software running on a DECstation 5000/200. The performance results were used to unco ver throughput and latency bottlenecks. We present a scheme for impro ving throughput when sending large messages by avoiding most checksum computations in a relatively safe manner. We also show that for the implementation we studied, reducing latency (when sending small messages) is a more dif ficult problem because processing overheads are spread over many operations; gaining a significant savings would require the optimization of many different mechanisms. This is especially important because, when processing a realistic workload, we ha ve found that non-data-touching operations

This paper describes the design and evaluates the performance of the DASH MP system, with an emphasis on its use of remapping. In addition, we study the system performance at a `microscopic' level by breaking down a user-level MP operation into short segments and measuring their individual performance. The results are used to evaluate and refine our design decisions

This paper presents the semi-passive replication technique -- a variant of passive replication -- that can be implemented in the asynchronous system model without requiring a membership service to agree on a primary. Passive

Continuous media playback suffers when a station 's operating system offers insufficient I/O throughput. Conventional I/O system structures support a memory--oriented read and write interface requiring the execution of user--level processes to facilitate playback, and can incur throughput degradation due to unnecessary data copies. Our splice mechanism supports a peer--to--peer model of I/O where a requesting application can associate a data source with its corresponding data sink, allowing for system optimizations in the data path implementation. In an experiment designed to simulate remote video playback, we present measurements indicating that use of our techniques resulted in a 55% gain in throughput as compared with conventional systems. 1 Introduction Computer and networking technology are beginning to offer a digital alternative to current analog cable and broadcast systems. Transmission and display of continuous media information including video is presently achievable with c...