The design of a hybrid measurement tool (LTRAMS) for a loosely-coupled multiprocessor (hypercube) is based on a previous design of a hybrid tool (TRAMS) which acquires event trace measurements on a tightly coupled sha...
详细信息
The design of a hybrid measurement tool (LTRAMS) for a loosely-coupled multiprocessor (hypercube) is based on a previous design of a hybrid tool (TRAMS) which acquires event trace measurements on a tightly coupled shared-memory MIMD architecture. A number of issues must be considered when applying this type of measurement tool to loosely-coupled, distributed-memory machines because of their architectural differences. These issues include synchronization and centralized vs. distributed functions. The LTRAMS triggering is done in software, while most of the measurement data sampling is done in hardware.
This tutorial will provide insight into the causes of the discrepancy between the raw transmission rate of physical media and the throughput rates observed by the end user. Some of the specific light-weight protocols ...
详细信息
This tutorial will provide insight into the causes of the discrepancy between the raw transmission rate of physical media and the throughput rates observed by the end user. Some of the specific light-weight protocols that are currently being developed for the network and transport layers of the OSI model will be examined. Finally, other measures of network performance, such as latency, which may actually be more important than throughput for real-time applications, will be discussed.
To assess the performance of a complete distributed system, the display of the locally obtained measurement results is generally not sufficient. New ways of evaluating, representing and displaying measurement data at ...
详细信息
To assess the performance of a complete distributed system, the display of the locally obtained measurement results is generally not sufficient. New ways of evaluating, representing and displaying measurement data at the central station are demanded. To cope with the vast amount of data collected by local monitors, the measurement information has to be preprocessed and filtered. To support performance debugging, the measurement data have to be presented at different levels: the whole system, selected nodes, clusters of program modules, single modules etc. At different levels different representations are recommended. Since application programmers are often not familiar with or not interested in system details, the measurement information has to be presented in an application-oriented manner. The applicability and practicability of high-level measurement support depends largely on its integration into the software development process.
A class of stochastic Petri nets with product form solution for the equilibrium state probabilities was described by A.A. Lazar and T.G. Robertazzi. For this class the Petri network consists of a number of linear sequ...
详细信息
A class of stochastic Petri nets with product form solution for the equilibrium state probabilities was described by A.A. Lazar and T.G. Robertazzi. For this class the Petri network consists of a number of linear sequences of places or subtasks with a vector of coordinates indicating which places are active. This work extends this class to the case where service times consist of a series of distinct exponential stages. Illustrative examples are given.
The author considers the Threshold Fastest Emptying policy. In the context of a system with saturated readers the TFE policy specifies: Serve M readers utilizing all processors until the queue for writers reaches some...
详细信息
The author considers the Threshold Fastest Emptying policy. In the context of a system with saturated readers the TFE policy specifies: Serve M readers utilizing all processors until the queue for writers reaches some threshold K. Empty the system from readers as quickly as possible by finishing serving the active readers (a non-preemptive policy). Then serve all available writers. The cycle repeats.
The response time is the most visible performance index to users of computersystems. End-users see individual response times, not the average. Therefore the distribution of response times is important in performance ...
详细信息
The response time is the most visible performance index to users of computersystems. End-users see individual response times, not the average. Therefore the distribution of response times is important in performance evaluation and capacity planning studies. However, the analytic results cannot be obtained in practical cases. A new queueing method is proposed to approximate the response-time distribution. Unlike the previous methods the proposed one takes into account the service-time distributions and routing behaviour. The reported results indicate that the method provides reasonable approximations in many cases.
In order to cope with the structure of (distributed systems) an event-driven monitor, called ZM4, has been developed. It provides an open number of distributed monitors called monitor agents. The monitor agents may be...
详细信息
In order to cope with the structure of (distributed systems) an event-driven monitor, called ZM4, has been developed. It provides an open number of distributed monitors called monitor agents. The monitor agents may be spatially distributed or locally concentrated. Data collection and buffering, however, is done independently in each monitor agent. In the author's project, the most interesting activities were the execution of processes which are modeled as a sequence of 'events of interest' and the execution of a collection of processes hence yields interleaved event sequences. Integrating monitoring and modeling is an essential for identifying events in a load to be monitored. The insight into an algorithm's dynamic behaviour on a given host architecture also helps to predict how the algorithm would be performed on other host architectures. a functional model supports monitoring, and the results of monitoring are used as attributes for the functional model, expanding it to a performance model.
This paper provides a simple unifying framework and concrete condition in order to conclude product forms results for interconnection or broadcasting protocol structures such as CSMA. Generally, such structures are re...
详细信息
This paper provides a simple unifying framework and concrete condition in order to conclude product forms results for interconnection or broadcasting protocol structures such as CSMA. Generally, such structures are referred to as random access schemes as scheduled transmissions can be blocked based upon the random state of the system, representing idle and busy transmitters or links. The actual access protocol is crucial for explicit expressions of the steady state distribution.
This paper is devoted to the analysis of waiting times in polling systems in which the stations are polled according to a general service order table. The polling system is modelled by a single-server multi-queue syst...
详细信息
This paper is devoted to the analysis of waiting times in polling systems in which the stations are polled according to a general service order table. The polling system is modelled by a single-server multi-queue system in discrete time, with non-zero switch-over times between the queues. The discrete-time formulation naturally fits the generally time-synchronized configuration of practical communication networks. Continuous-time results are easily obtained via a limiting procedure. The main result is an exact expression for a weighted sum of the mean waiting times at the various queues. It is obtained via an extension of the principle of work conservation.
As multiprocessors with large numbers of processors become more prevalent, we face the task of developing scheduling algorithms for the multiprogrammed use of such machines. We examine the quality of processor allocat...
详细信息
As multiprocessors with large numbers of processors become more prevalent, we face the task of developing scheduling algorithms for the multiprogrammed use of such machines. We examine the quality of processor allocation decisions under multiprogramming that can be made with several different high-level characterizations of application parallelism. We demonstrate that decisions based on parallelism characterizations with two to four parameters are superior to those based on single-parameter characterizations (such as fraction sequential or average parallelism). The results are based predominantly on simulation, with some guidance from a simple analytic model.
暂无评论