Ermittlung des Leistungsverhaltens von Timesharing-Systemen

In order to describe the performance of TS-systems continuous measurements of the performance-related quantities (controls variables, eV) are carried out. By using statistical methods the set of evs is ordered (multiple regression, clusteranalysis) or reduced (stepwise multiple regression) according to their impact on the user-specific performance measure 'system response time'. The thus selected evs. which are significant for the operation of the timesharing system. are used for the construction of a TS-benchmark (TS-BM). TS-BMs ensure the reproducibility of the workload. Theyare generated modularly from basic scripts BI" .. BM (M=3). They are used as system-drivers in the forms of quasi-. pseudoor dynamic benchmarks which consist of script-products (parallelity) or script-polynomials (seriality, to obtain average eV-values for a given operation cycle). Workload approximation is achieved through selection (S-polynomials) and through parameter tuning (think time. transaction rate. I/O-rate. epu time consumption) when S-products are used. Dynamic benchmarks consist of symmetric S-polynomials ~ or of se ries of homogeneous S-products Si' HMN (B 1 ,·· .• BM) = 1:i a i Si (B 1•··· ,BM)· The homogeneity of S-products guarantees the real workload's approximate simulation of the functional spectrum. It is shown that using dynamic benchmarking system performance diagrams, resource utilizations, aud response time behavior of existing systems' (IBM/370) workload can be modeled to measurement precision. Performance changes due to system alterations (hardware, software. system parameters) can be evaluated by measuring the maximum system performance, performance ratios. and productivity gains. Workload changes are recognized automaticallYi this leads to adjustments of the TS-BMs as weIl as to important insights as a basis for management decisions for application and system planning.