Maria Calzarossa and Giuseppe Serazzi. A characterization of the variation in time of workload arrival patterns. IEEE Trans. Comput., C-34(2):156--162, Feb 1985.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....0 4 8 12 16 20 24 submitted hour time of day weekdays weekends 0 1 2 3 4 5 6 10 12 14 16 18 arrival rate time of day Fig. 3. Left: cyclic job arrival pattern in the NASA Ames iPSC 860 (from [10] Right: model of Calzarossa and Serazzi for workload at the University of Pavia [2]. submitted strongly correlates with the time of day and the day of the week (Fig. 3) However, the simpler continuous model is the one that system performance evaluation almost always uses in practice. This has the unfortunate effect of excluding the evaluation of scheduler optimizations that ....

....possible to increase the load by uniformly reducing the interarrival times, because such a practice will also shrink the cycle length. Regrettably, very little work has been done on the derivation of realistic models. The only detailed model we know of was proposed by Calzarossa and Serazzi [2]. This model uses a polynomial of degree 8 to model the changing arrival rate of interactive work (Fig. 3) The proposed polynomial for normal days is (t) 3:1 Gamma8:5t 24:7t 7 8 where (t) is the arrival rate at time t, and t is in the range [ Gamma0:5: 0:5] and should be ....

M. Calzarossa and G. Serazzi, "A characterization of the variation in time of workload arrival patterns". IEEE Trans. Comput. C-34(2), pp. 156--162, Feb 1985.

....in recent years. Maybe the most prominent are the study of Internet traf c patterns [52, 62, 75] and world wide web trac patterns, with the intent of using the knowledge to evaluate server performance and caching schemes [5, 18, 6] Other examples include studies of process arrivals and runtimes [12, 37], le systems [36] and video streams [48] In the area of parallel systems, descriptive studies of workloads have only started to appear in recent years [29, 76, 58, 27, 14] There are also some attempts at modeling [10, 28, 21, 41, 23, 54, 15] and on line characterization [34] But where does ....

M. Calzarossa and G. Serazzi, \A characterization of the variation in time of workload arrival patterns". IEEE Trans. Comput. C-34(2), pp. 156-162, Feb 1985.

....300 400 500 600 700 SDSC of Jobs Figure 10 Histograms of arrival hour for our reference workloads Methodology As Figure 10 suggests, it is very hard to model the job arrival time through common distributions. In this work, we apply the methodology proposed by Calzarossa and Serazzi [17] to numerically fit a polynomial to the job arrival a . Although the methodology was conceived to model the process arrival for a uniprocessor time shared system, we found it to be applicable in our scenario, namely job arrivals for parallel supercomputers. In order to fit a polynomial to the ....

....applicable in our scenario, namely job arrivals for parallel supercomputers. In order to fit a polynomial to the arrival rate found in our reference workload logs, we must derive the arrival rate from the arrival instants (which are in the log) Following the methodology of Calzarossa and Serazzi [17], we smoothed out the arrival rate by using a moving average estimator. More precisely, for a given time m (in minutes) the arrival a (m) is estimated using all arrivals in the 10 minute interval cen tered on m. Also, in order to avoid numerical instabilities while fitting the data, we scale a ....

[Article contains additional citation context not shown here]

M. Calzarossa and G. Serazzi. A characterization of the variation in time of workload arrival patterns. IEEE Transactions on Computers, vol. C-34, (no.2), Feb. 1985. p.156-62.

....for evaluating the metacomputer scheduling techniques. 4.1 Job Arrivals As in [18] we use an open on line model where jobs are continuously fed into the system while the scheduler is working. The rate at which new jobs enter the metacomputer is modeled by a polynomial of degree 8 as described in [7]. Thus, the arrival rate at any given time t is (t) 3:1 Gamma8:5t 24:7t 2 130:8t 3 107:7t 4 804:2t 5 2038:5t 6 1856:8t 7 4618:6t 8 (2) with Gamma0:5 t 0:5 representing the time span from 8:30 AM to 6:00 PM. During the remaining time, we estimate an arrival rate of one job ....

M. Calzarossa and G. Serazzi. A characterization of the variation in time of workload arrival patterns. IEEE Transactions on Computers, Vol. C-34, :2, 156-162, 1985.

....will be effective if the requirements are known a priori, which implies prediction. Prediction of resource requirements is all the more relevant in systems which exhibit dynamic workload characteristics. Several models were proposed to describe the workload characteristics of applications [1, 2, 3, 4, 5, 6]. Early researchers have proposed a string based representation of the state of the networked computing environment [7] In fact, the inherently hierarchical nature of such systems has also been captured using such an approach in the form of Context Free Grammars [8, 9] The most recent of these ....

....units in 64kbps Time in Seconds VP 1 11 in 11 Node Network Bandwidth Allocation through SSR Model Actual Demand Fig. 6. VP Bandwidth Allocation for the 4 Node and 11 Node Networks through the SSR Model goal of any workload modeler. Attempts such as dynamic models [3] and time based models [4] are endeavors in this direction. The primary concern of a modeler working with learning and adaptive models is to quantify the ability of the model to learn, thereby quantifying the intelligence associated with the model. In this work, we have addressed these concerns as well as successfully ....

Calzarossa M and Serazzi G. A Characterization of the Variation in Time of Workload Arrival Patterns. IEEE Trans. on Computers, C-34(2):156--162, 1985.

....requirements, runtime, I O needs, etc. Unfortunately, there is not much hard data that has been measured about typical internal structural distributions. Modelling Job Arrivals and Internal Job Structure are active areas of research. For detailed information about both the reader should read [1, 8, 15, 9, 10, 12, 6]. The internal job characteristics supported by the simulator are: ffl Degree of parallelism ffl Initial Processor Arrival ffl I O Distribution ffl Communication and Computation Distributions ffl Synchronization Distribution ffl Task Execution Distribution ffl Initial Task Distribution ffl ....

M. Calzarossa and G. Serazzi. A characterization of the variation in time of workload arrival patterns. IEEE Transactions on Computers, 34(2):156--162, February 1985.

....system support jobs are not) On weekdays, a daily cycle is obvious, with a high submission rate during the working day and a low rate at night. The peak is in the late morning, with a noticeable drop during lunch. This pattern is similar to known results from uniprocessor interactive systems [4]. The peak value is about one job every 2.6 minutes, on average. The implication is that the load on the system does not change too often, so scheduling schemes can afford to optimize the way resources are shared. The gradual rise in activity before 8 AM is attributed to users from the East Coast ....

M. Calzarossa and G. Serazzi, "A characterization of the variation in time of workload arrival patterns". IEEE Trans. Comput. C-34(2), pp. 156--162, Feb 1985.

....evaluated. 7 Related and Future Work The suggestion that workload modeling should be based on measurements is not new [15, 1] However, relatively little work has been done on this subject. Examples include work on characterizing the distribution of durations [26, 17] and on the arrival process [3], both in the context of Unix workstations. In the area of parallel systems, descriptive studies of workloads have only started to appear in recent years [12, 38, 33, 9] There are also some attempts at modeling [2, 7, 20] and on line characterization [16] Practically every section of this paper ....

M. Calzarossa and G. Serazzi, "A characterization of the variation in time of workload arrival patterns". IEEE Trans. Comput. C-34(2), pp. 156--162, Feb 1985.

....simulates the generation of competing processes, running in parallel on interactively used workstation clusters, and taking away CPU cycles there. Processes can be generated by using xed arrival and departure rates, variable arrival and departure rates provided by trace les (Calzarossa, Serazzi [3]) trace les of real processes (Zhou [17] and user behavior graphs (Calzarossa, Serazzi [4] By constructing stochastic models of real parallel applications or running real applications, di erent load balancing schemes can be simulated and compared with each other. The structure of the whole ....

Calzarossa M., Serazzi G., A Characterization of the Variation in Time of Workload Arrival Patterns, IEEE Transactions on Computers C-34-2 (1985), pp. 156-162.

....queries of the test application take place according to an exponential distribution. Figure 4: Load distribution The trend of the distribution of the load on the system shown in Fig. 4 was obtained by discretizing the load curve of an academic environment obtained by Calzarossa and Serazzi in [2]. The rate of generation of queries in the different hours was obtained by making the load on the system to result equal to 100 in the hour of largest use (between 11:00 and 12:00) and by proportionally reducing such value in the other hours, according to the scale factor obtained from Fig. 4. ....

M. Calzarossa and G. Serazzi. A Characterization of the Variation in Time of Workload Arrival Patterns. IEEE Transactions on Computers, C-34(2):156--162, February 1985.

....of runtimes than estimates that do not use historical knowledge. The lower the CV, the more accurate the estimates are likely to be. We will focus on choosing what attributes to use to classify jobs. There have been only a few detailed workload characterization studies of parallel systems [CS85, PBK91]. Furthermore, the studies that do exist do not focus on different ways of classifying jobs, with the exception of Feitelson and Nitzberg s [FN95] analysis of the workload of the 128 node NASA Ames iPSC 860 hypercube. Feitelson and Nitzberg classify jobs by name, user, and number of processors, ....

M. Calzarossa and G. Serazzi. A characterization of the variation in time of workload arrival patterns. IEEE Transactions on Computers, C-34(2):156-- 162, February 1985.

....processes as well (Kunz [13] There are, however, various ways in modeling workstation workload. Among them are, for instance, Poisson processes (Allen [1] with constant arrival and departure rates (Kunz [13] Poisson processes with variable arrival rates during the day (Calzarossa, Serazzi [4] [6] Markovian type models (Haring [9] Calzarossa, Serazzi [5] and probabilistic context free grammars (Rhagavan, Joseph [15] In this paper, a simulator based on MISS PVM (Kvasnicka [14] is presented, which is designed to simulate different load balancing strategies for parallel programs ....

Calzarossa M., Serazzi G., A Characterization of the Variation in Time of Workload Arrival Patterns, IEEE Transactions on Computers C-34-2 (1985), pp. 156-162.

....of the I O time. They analyse the arrival rate of jobs, and find a common pattern of workload increasing from 8:00am to noon, remaining roughly constant until 5:00pm, and then decreasing until 9:00pm. An earlier study of job arrival patterns in a multiprocessor was done by Calzarossa and Serazzi [CS85] They analyse 14 one day periods using polynomial fitting techniques to derive functions for the arrival rate based on the time of day and cluster analysis to find groups of applications that have similar arrival behaviour. They verify their findings with a second set of data from a different ....

M. Calzarossa and G. Serazzi. A characterization of the variation in time of workload arrival patterns. IEEE Transactions on Computers, C-34(2):156-- 162, February 1985.

....workload characterization to the dispersion of resource demands of the clustered commands is also evaluated. The results obtained have shown that the clustering method for workload model design is reasonably accurate and the sensitivity is quite low. Numerical fitting techniques are used in [20] for the analysis of the fluctuations in the arrival patterns of the workload components and the construction of a parametric model able to provide a concise representation of the analyzed phenomena. The arrival times of the jobs at the system are collected during one month of operation. Some ....

..... Figure 4: Estimated (dotted curve) and polynomial (solid curve) arrival rate functions [20]. application of these methods to the various days has shown that an eight degree polynomial function is a suitable representation of all the analyzed arrival processes. Three representative patterns with similar behaviors have been identified in the data by means of the clustering applied to the ....

M. Calzarossa and G. Serazzi. A Characterization of the Variation in Time of Workload Arrival Patterns. IEEE Trans. on Computers, C-34(2):156--162, 1985.

No context found.

Maria Calzarossa and Giuseppe Serazzi. A characterization of the variation in time of workload arrival patterns. IEEE Trans. Comput., C-34(2):156--162, Feb 1985.

No context found.

M. Calzarossa and G. Serazzi. A characterization of the variation in time of workload arrival patterns. IEEE Transactions on Computers, vol. C-34, (no.2), Feb. 1985. p.156-62.

No context found.

M. Calzarossa and G. Serazzi. A Characterization of the Variation in Time of Workload Arrival Patterns. IEEE Trans. on Computers, C-34(2)) 1985.

No context found.

M. Calzarossa and G. Serazzi. A characterization of the variation in time of workload arrival patterns. IEEE Transactions on Computers, C-34(2):156--162, February 1985.

No context found.

M. Calzarossa and G. Serazzi. A characterization of the variation in time of workload arrival patterns. IEEE Transactions on Computers, Vol. C-34, :2, 156-162, 1985.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC