Citation Context

... modern implementations of the TCP/IP protocol include a dynamic predictor of end-toend round-trip time based on stochastic gradient filter [28]. We follow the exposition of the technique provided in =-=[22]-=- which includes a description of a very efficient implementation for the Unix kernel. We define " +'&4 1 &4 " *s 1 +(& 8)& $3&()s (10) for a gain 0 #*&+# 1 ....

Citation Context

...he NWS influences both the implementation of the overall system and the forecasting techniques we have chosen. Since the problems of non-intrusive resource monitoring [26, 21, 9] and load forecasting =-=[24, 3, 18, 23, 8]-=- both pose open research questions, we have separated the sensory and forecasting functions of the NWS. The resulting modular design is intended to provide a general facility in which a variety of dif...

Citation Context

...he NWS influences both the implementation of the overall system and the forecasting techniques we have chosen. Since the problems of non-intrusive resource monitoring [26, 21, 9] and load forecasting =-=[24, 3, 18, 23, 8]-=- both pose open research questions, we have separated the sensory and forecasting functions of the NWS. The resulting modular design is intended to provide a general facility in which a variety of dif...

Citation Context

...mum. In the experiments presented in the next section, we set 5 != K != 50. Stochastic gradient or recursive prediction error estimators are powerful predictive techniques with recursive formulations =-=[25]-=-. For example, modern implementations of the TCP/IP protocol include a dynamic predictor of end-toend round-trip time based on stochastic gradient filter [28]. We follow the exposition of the techniqu...

Citation Context

...he NWS influences both the implementation of the overall system and the forecasting techniques we have chosen. Since the problems of non-intrusive resource monitoring [26, 21, 9] and load forecasting =-=[24, 3, 18, 23, 8]-=- both pose open research questions, we have separated the sensory and forecasting functions of the NWS. The resulting modular design is intended to provide a general facility in which a variety of dif...

Citation Context

...number of samples to consider in the window. Note that for K = 0, SW AV G uses the last measurement only as a predictor. That is, LAST (t) = SW AV G(t; 0) (7) Recent work by Harchol-Balter and Downey =-=[20]-=- indicates that this is a useful predictor for CPU resources, hence we include it as a separate method. The choice of K for SW AV G may be difficult to determinesa priori for each resource, and in fac...

Citation Context

...en define MIN MSE(t) = predictor f (t) (17) 2 The autoregressive model is applicable if the decay in the autocorrelation function is exponential and the value of p is set to the duration of the decay =-=[16]-=-. Our current implementation of the NWS does not attempt to determine the suitability of AR for a particular resource. Instead, it assumes that the autoregressive model is applicable, and tracks the p...

Citation Context

...uare error and mean percentage error respectively. 5.3. Summary and Analysis of Network Forecasting Performance A more detailed description and analysis of the forecasting performance can be found in =-=[31]-=-. Tables 3 and 4 summarize the performance of the best predictors for each setting. Notice that GRAD is the overall best predictor of throughput if mean square error is used as an accuracy measure. It...

Citation Context

... is the autocorrelation function for the series of N measurements taken. The Levinson Recursion requires a set of partial correlation (PARCOR) coefficients which can also be derived recursively. Burg =-=[6]-=- and more recently Haddad and Parsons [19] describe a recursive algorithm for calculating both the PARCOR and autoregression coefficients from which we derive our current implementation. We omit the d...

Citation Context

... applications using dynamic performance forecasting as the basis for scheduling, and other work indicates that dynamic information can be used to enhance the performance of world-wideweb applications =-=[7]-=-. In this paper, we focus on the problem of network performance forecasting to support parallel application scheduling. We have developed a prototype of the NWS that forecasts network performance (lat...

Citation Context

...chniques which require a computationallyintensive "fitting" phase. The ARIMA models described in [3], the self-similarity analysis outlined in [24], and the semi-nonparametric techniques dis=-=cussed in [13, 14]-=-, all provide immediately promising avenues of investigation. We would like to discern the relationship between the computational complexity devoted to making a forecast its accuracy. We also plan to ...

Citation Context

...d to limit the intrusiveness of the NWS influences both the implementation of the overall system and the forecasting techniques we have chosen. Since the problems of non-intrusive resource monitoring =-=[26, 21, 9]-=- and load forecasting [24, 3, 18, 23, 8] both pose open research questions, we have separated the sensory and forecasting functions of the NWS. The resulting modular design is intended to provide a ge...

Citation Context

...ts Agency/ITO, Distributed Object Computation Testbed, ARPA order No. D570, Issued by ESC/ENS under contract #F19628-96-C-0020. y email: rich@cs.ucsd.edu, http://www-cse.ucsd.edu/users/rich/ policies =-=[5, 11, 15, 30]-=-. To gain the desired levels of performance, however, the scheduling methods described in these works depend on predictions of the performance deliverable to the application from the available metacom...

Citation Context

...t. Systems such as those outlined in [15, 30, 29, 10] can use NWS forecasts to parameterize their respective scheduling methods and thereby generate schedules that are sensitive to load variation. In =-=[5, 4]-=- we report on the efficacy this technique for parallel applications in production distributed computing environments. As a result of its success, we are currently implementing versions for two product...

Citation Context

...can also serve as a useful predictor, particularly if the measurement sequence contains randomlyoccurring, asymmetric outliers. Our presentation of these techniques follows the exposition in [19] and =-=[12]-=-. The median over a sliding window of fixed length whose leading edge is the most recent measurement is used as the forecast for the next measurement. That is, we define SortK = the sorted sequence of...

Citation Context

...recasting, and as such, term the service the Network Weather Service (NWS). We have developed the NWS for use by schedulers in a networked computational environment. Systems such as those outlined in =-=[15, 30, 29, 10]-=- can use NWS forecasts to parameterize their respective scheduling methods and thereby generate schedules that are sensitive to load variation. In [5, 4] we report on the efficacy this technique for p...

Citation Context

...d to limit the intrusiveness of the NWS influences both the implementation of the overall system and the forecasting techniques we have chosen. Since the problems of non-intrusive resource monitoring =-=[26, 21, 9]-=- and load forecasting [24, 3, 18, 23, 8] both pose open research questions, we have separated the sensory and forecasting functions of the NWS. The resulting modular design is intended to provide a ge...

Citation Context

...ity to predict the TCP/IP end-to-end throughput and latency that is attainable by an application using systems located at different sites. Such network forecasts are needed both to support scheduling =-=[5]-=-, and by the metacomputing software infrastructure to develop quality-of-service guarantees [10, 17]. Keywords: scheduling, metacomputing, quality-ofservice, statistical forecasting, network performan...

Citation Context

...d to limit the intrusiveness of the NWS influences both the implementation of the overall system and the forecasting techniques we have chosen. Since the problems of non-intrusive resource monitoring =-=[26, 21, 9]-=- and load forecasting [24, 3, 18, 23, 8] both pose open research questions, we have separated the sensory and forecasting functions of the NWS. The resulting modular design is intended to provide a ge...

Citation Context

...he NWS influences both the implementation of the overall system and the forecasting techniques we have chosen. Since the problems of non-intrusive resource monitoring [26, 21, 9] and load forecasting =-=[24, 3, 18, 23, 8]-=- both pose open research questions, we have separated the sensory and forecasting functions of the NWS. The resulting modular design is intended to provide a general facility in which a variety of dif...

Citation Context

...n using systems located at different sites. Such network forecasts are needed both to support scheduling [5], and by the metacomputing software infrastructure to develop quality-of-service guarantees =-=[10, 17]-=-. Keywords: scheduling, metacomputing, quality-ofservice, statistical forecasting, network performance monitoring 1. Introduction As network technology advances, the resulting improvements in interpro...

Citation Context

...chniques which require a computationallyintensive "fitting" phase. The ARIMA models described in [3], the self-similarity analysis outlined in [24], and the semi-nonparametric techniques dis=-=cussed in [13, 14]-=-, all provide immediately promising avenues of investigation. We would like to discern the relationship between the computational complexity devoted to making a forecast its accuracy. We also plan to ...

Citation Context

...ve techniques with recursive formulations [25]. For example, modern implementations of the TCP/IP protocol include a dynamic predictor of end-toend round-trip time based on stochastic gradient filter =-=[28]-=-. We follow the exposition of the technique provided in [22] which includes a description of a very efficient implementation for the Unix kernel. We define GRAD(t; g) = (1 \Gamma g)sGRAD(t \Gamma 1; g...

Citation Context

...ts Agency/ITO, Distributed Object Computation Testbed, ARPA order No. D570, Issued by ESC/ENS under contract #F19628-96-C-0020. y email: rich@cs.ucsd.edu, http://www-cse.ucsd.edu/users/rich/ policies =-=[5, 11, 15, 30]-=-. To gain the desired levels of performance, however, the scheduling methods described in these works depend on predictions of the performance deliverable to the application from the available metacom...

Citation Context

...he NWS influences both the implementation of the overall system and the forecasting techniques we have chosen. Since the problems of non-intrusive resource monitoring [26, 21, 9] and load forecasting =-=[24, 3, 18, 23, 8]-=- both pose open research questions, we have separated the sensory and forecasting functions of the NWS. The resulting modular design is intended to provide a general facility in which a variety of dif...

Citation Context

...ts Agency/ITO, Distributed Object Computation Testbed, ARPA order No. D570, Issued by ESC/ENS under contract #F19628-96-C-0020. y email: rich@cs.ucsd.edu, http://www-cse.ucsd.edu/users/rich/ policies =-=[5, 11, 15, 30]-=-. To gain the desired levels of performance, however, the scheduling methods described in these works depend on predictions of the performance deliverable to the application from the available metacom...

Citation Context

...n using systems located at different sites. Such network forecasts are needed both to support scheduling [5], and by the metacomputing software infrastructure to develop quality-of-service guarantees =-=[10, 17]-=-. Keywords: scheduling, metacomputing, quality-ofservice, statistical forecasting, network performance monitoring 1. Introduction As network technology advances, the resulting improvements in interpro...

Citation Context

... modern implementations of the TCP/IP protocol include a dynamic predictor of end-toend round-trip time based on stochastic gradient filter [28]. We follow the exposition of the technique provided in =-=[22]-=- which includes a description of a very efficient implementation for the Unix kernel. We define GRAD(t; g) = (1 \Gamma g)sGRAD(t \Gamma 1; g) + gsvalue(t) (10) for a gain (0 ! g ! 1). The choice of g ...