P. Mehra, Automated Learning of Load Balancing Strategies for a Distributed Computer System, Technical Report, University of Illinois, 1993  Home/Search   Document Details and Download   Summary   Related Articles   Check

....user space. This user space should contains a set of different resources on different machines allowing the user to satisfy 80 of his service requests. Therefore at first a characteristic of typically used processes and programs of the respective users must be created, stored and considered [11][16] The use of a small, fixed set of user classes as suggested in [16] is not proper for doing so, because of the resource needs of the users may differ to much. For doing so we suggest the application of statistic methods to characterize the processes typically started by the respective user, ....

P. Mehra, Automated Learning of Load Balancing Strategies for a Distributed Computer System, Technical Report, University of Illinois, 1993

....increase by introducing load balancing Once the necessary time keys from each host have been obtained, a policy for how to place the servers in the recursion tree has to be determined. 5.5. 1 Strategies Load balancing strategies can be classified after the following criteria as referred in [Mehra 93] 9 The nice value of a process is telling the kernel what the priority this process shall have. A high nice value means low priority. 5.5 Performance increase by introducing load balancing 65 1. Whether processes once placed at a host can be migrated to other hosts if the host becomes heavily ....

Pankaj Mehra (1993): "Automated learning of load-balancing strategies for a distributed computer system." PhD Thesis, University of Illinois at Urbana-Champaign, 1993, chapter 1. [ftp://ftp.ibr.cs.tubs. de/pub/docs/DistributedSystems/Loadbalance/Mehra/Thesis.ps.gz].

....has worked well for some years, we are aware that it has a couple of shortcomings and we would like to improve it. We conclude by listing some problems for future work. We believe that answers to theses questions will lead to such an improvement. 1. avenrun[0] has been criticized (e.g. [12]) as being not appropriate to compare workload in a heterogeneous environment. We choose it because it can be obtained easily and without super user privileges. 2. A number of existing systems (e.g. shortest expected delay [11] or CoCheck [9] uses a load index of ff Delta (1 avenrun) This ....

....KIV System: Systematic Construction of Verified Software. In Proc. CADE 11, pp. 753 757, Springer, 1992. 11] B. Schnor, S. Petri, R. Oleyniczak, and H. Langendorfer. Scheduling of Parallel Applications on Heterogeneous Workstation Clusters. In Proc. PDCS 96, vol. 1, pp. 330 337, ISCA, 1996. [12] S. Stille. Lastbalancierung in verteilten Systemen. Masters thesis, Braunschweig University of Technology, IBR, 1993. 13] C. Weidenbach, B. Gaede, and G. Rock. SPASS FLOTTER, Version 0.42. In Proc. CADE 13, pp. 141 145, Springer, 1996. 22 23 24 25 26 27 28 29 30 31 A ....

P. Mehra, "Automated Learning of Load Balancing Strategies for a Distributed Computer System", Technical Report, University of Illinois, 1993.

....can be registered online only with statistical means or off line with so called work load generators. For realization, the following adaptive strategies are proposed to get a basis of the necessary performance prediction: a registration and learning of concrete load situations as described in (Mehra 1993), Wilhelms 1994) using average CPU time profiles of the processes to predict and to use (a) the future load (Goswami, Devarakonda and Iyer 1993) without calculating the already done work of a process or (b) the time, when the machine will be idle again; a calculation of stochastic dependency ....

....done work of a process or (b) the time, when the machine will be idle again; a calculation of stochastic dependency parameters between processes. Because of the simplicity we will follow the approach (2b) Furthermore we want to use in contrast to some other systems (see Epema et al. 1995) and (Mehra 1993)) online algorithms (as suggested in (Goswami, Devarakonda and Iyer 1993) in order to enable a reaction to changes in the system during its work (e.g. changing execution times of programs processes because of new values of calling parameters or input data from the environment) The tasks of ....

P. Mehra. 1993. "Automated Learning of Load Balancing Strategies for a Distributed Computer System".

....agents. This is in sharp contrast to our work. In addition, there are differences between the type of model we use and the model presented in the above mentioned work and in other work on distributed computer systems. Applications of learning algorithms to load balancing problems are given by Mehra (1992), Mehra and Wah (1993) However, in that work as well, the agents (sites, in the authors terminology) have the ability to communicate and to exchange workload values, even though such values are subject to uncertainty due to delays. In addition, differently from our work, the learning activity is ....

Mehra, P. (1992). Automated Learning of Load-Balancing Strategies For A Distributed Computer System. Ph.D. thesis, Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign.

....arriving activation trees. In [68] the authors present location policies that adapt to system load to avoid instability. Mehra and Wah [52] use comparator neural networks to predict current workload indices on remote hosts using outdated information about resource utilization. Mehra s thesis [51] describes a whole load balancing system based on automated strategy learning using comparator neural networks. We have also found some success with a neural networks approach. Indeed, it is interesting to note that machine learning [53] approaches are also finding success in related scheduling ....

MEHRA, P. Automated Learning of Load-Balancing Strategies for a Distributed Computer System. PhD thesis, University of Illinois at Urbana-Champaign, 1993.

....CPU cycles. The combined power of interconnected workstations can exceed that of mainframes if we can make them work together and distribute the computation load among them evenly. Beside initial placement, process migration is a means for load balancing and utilizing otherwise idle machines [6, 14, 17, 28, 18]. It can also improve fault tolerance, by evacuating hosts prior to regular shutdown, or through checkpointing [14, 18] A variety of process and task migration systems have been developed. All these systems have limitations that make them usable only for small classes of applications or in ....

Pankaj Mehra. Automated Learning of Load-Balancing Strategies for a Distributed Computer System. PhD thesis, University of Illinois at Urbana-Champaign, 1993.

....workload measures, which indicate each site s load; and decision policies, which determine both the conditions under which tasks are migrated and the destinations of incoming tasks. In this paper, we address automated learning of load measures; learning of policy parameters is described elsewhere [8]. Figure 1 shows the parameterized policy considered in this paper. The sender side rules (SSRs) are evaluated at s, the site of arrival of a task. Reference can be either 0 or MinLoad; the other parameters d, q 1 , and q 2 take non negative real values. A remote destination, r, is picked ....

....them as often as desired. Representative test jobs must be run under such synthetic workloads, and the utilization of all resources recorded periodically. Both utilization levels and completion times must be measured with low overhead and high precision. With these in mind, we developed DWG [8, 10], a synthetic workload generator that helps us build our database of decision points. DWG can measure and control the utilization levels of four key resources: CPU, memory, disk, and network. It supports a variety of data collection operations: i) precise measurement of resource utilization ....

[Article contains additional citation context not shown here]

P. Mehra, Automated Learning of Load Balancing Strategies for a Distributed Computer System, Ph.D. Thesis, Dept. of Computer Science, Univ. of Illinois, Urbana, IL, Dec. 1992.

....patterns approximate those from real workloads. In order to achieve high quality generation, we need to first parameterize the generation mechanisms, then experiment with many different parameter sets on a prototype system, and finally select the best one for our load balancing experiments. [9]. Our experiments were carried out on a configurationally heterogeneous system consisting of (i) a diskless Sun 3 50 with 4 Mbytes of RAM; ii) a diskful Sun 3 50 with 4 Mbytes; iii) a diskful Sun 3 260 with 8 Mbytes; and (iv) a diskless Sun 3 60 with 24 Mbytes. The four workstations were ....

....of UNIX sort utility with different file sizes and memory requirements, two jobs of the UNIX uncompress program, and two of the Perfect Club benchmarks. 3] More detailed evaluation using other benchmarks (including other perfect club benchmarks) can be found in the first author s Ph.D. thesis. [9] The choice of benchmarks was giverned by our desire to achieve a job mix with sufficient variety so that the load index would be forced to take into account resources other than the CPU. For instance, sorting with large amount of memory and compressing large files achieve, respectively, pure ....

[Article contains additional citation context not shown here]

P. Mehra, Automated Learning of Load Balancing Strategies for a Distributed Computer System, Ph.D. Thesis, Dept. of Coputer Science, Univ. of Illinois, Urbana, IL, Dec. 1992.

....problem in the evaluation of load balancing strategies, the case with no load balancing is often used as a point of reference. This approach requires that two sets of experiments be performed using exactly the same set of jobs and loading conditions: once with, and once without, load balancing [38]. To illustrate Issue 5, note that completion time is not a well defined function of status variables alone, and job descriptors are rarely available when decisions are made. The completion time based objective functions are ill defined and can be evaluated only over state sequences. When queuing ....

....of confidence sought is just enough to allow the learning system to minimize the risk of choosing a bad strategy due to insufficient testing. Population based learning has been used to learn strategies for static load balancing with dependent jobs [92] dynamic load balancing with independent jobs [38], VLSI test generation [140] and stereo vision [141] Table 9 summarizes the characteristic features of the four architectural models. Table 10 reviews their applicability to the strategy learning problems described at the end of Section 1. 5 Conclusions A strategy learning problem is a triple ....

P. Mehra, Automated Learning of Load Balancing Strategies for a Distributed Computer System. Urbana, IL: Ph.D. Thesis, Dept. of Computer Science, Univ. of Illinois, Dec. 1992.

....experiments, the design of workload measures that indicate each site s load, and the design of load balancing policies that determine both the conditions under which tasks are migrated and the destinations of incoming tasks. In this paper, we summarize the design of the first two components [8, 9] and present a population based learning algorithm for tuning policy parameters. We describe the architecture of SMALL (Systematic Method for Automated Learning of Load balancing policies) our learning system, in Section 2. Section 3 presents the various phases of SMALL s operation: data ....

....the formula shown for Phase IV. In the remaining sections of this paper, we present our approach to the automated tuning of load balancing policy parameters using integrated (population based and point based) learning. Details for the other components are not included due to space limitation [8, 9]. 4. INTEGRATED LEARNING OF LOADBALANCING POLICIES 4.1. The Policy Learning Problem Figure 1 shows the sender side and receiver side rules for the load balancing policies considered in this paper. There is one set of rules per site. The rules are parameterized; the number of parameters depends ....

P. Mehra, Automated Learning of Load Balancing Strategies for a Distributed Computer System, Ph.D. Thesis, Dept. of Computer Science, Univ. of Illinois, Urbana, IL, Dec. 1992.

No context found.

P. Mehra, "Automated Learning of Load Balancing Strategies for a Distributed Computer System", Technical Report, University of Illinois, 1993.

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