Simulating the architecture Sample Clauses

Simulating the architecture. In order to the compare the performance of a fine grained description of available resources regarding different computation tasks, and information about job prefer- ences, we model our systems as a queuing network composed of 5 nodes, correspond- ing to our heterogeneous test bed, plus a scheduler which dispatches arriving jobs to the resources. In the global scheduler depicted in Figure 1, different scheduling xxxxxx- xxxx can be used, e.g., a round xxxxx job allocation. However, for the performance- based SLA architecture we favor the usage of the Congruent Policy job allocation, which takes into account the appropriate resource properties. Moreover, we considered two more job allocation strategies based on information derived using the established ISO and HPL benchmarks respectively. Our objective is to minimize the Response Time of the system, that takes into account the time that a job takes to be executed (service time) plus the time spent in queue (waiting to be executed). In the simulation we considered a workload composed of two parallel applications (linear algebra and isosurface extraction) that have been modelled as two open classes with exponentially distributed inter-arrival and service times [19]. Service times are obtained through a real experimentation on the base of the benchmark values as re- ported in Table 2. They can be considered as the results of the prediction component. Table 2. Mean service times of each application class. (in parentheses the number of processors spawned for each resource) IBM (32) Xxxxxxxxxxxx (32) SC1458 (128) Paperoga (8) Cluster1 (16) ISO 2.4 3 35 13 7 HPL 33 25 4.5 55 62 In Figure 5 the response times of each strategy at increasing workloads are shown. It is immediately clear that the proposed performance-based SLA outperforms the other schedulers. This is not surprising since each resource is exploited as its best re- spect to the incoming workloads, i.e. each application is allocated to the resources that execute the code in the most efficient way, in our analysis with minor execution time. It leads to faster execution and lower waiting time. Both parameters impact (in this case positively) on the response time. An increase of computation intensive workloads also influences our scheduling mechanism, however the growth of response time is moderate compared with other tested strategies.