November 25, 2009 Conference paper Open Access

Simulations of distributed systems in a computing centre

Dongiovanni, Danilo; Ronchieri, Elisabetta; Dal Pra, Stefano; dell'Agnello, Luca; Ferrari, Tiziana; Antonelli, Stefano; Cavalli, Alessandro; Gregori, Daniele; Martelli, Barbara; Prosperini, Andrea; Ricci, Pier Paolo

Modern computing centres addressed to High Energy Physics user communities have to deal with rapidly hard- ware and software systems evolution. These centres normally face a variety of problems associated with the dimensioning and configuration of several services which must satisfy per- formance targets under different usage patterns. Therefore, the identification of key variables and the estimation of their impact on services performances is challenging. For example, given an hardware-software configuration for a considered service, how will service performance vary in relation to user dependent settings? Will it be able to support a certain number of requests per minute over the common parameter ranges? In addition, it is difficult to generalize the impact of same settings over different usage scenarios. Therefore, the design of a mathematical model able to relate services performance to key variables in the user computing patterns under common hardware-software settings can help to optimize the exploitation of computing resources.  In the present work, starting from the analysis of a typical job of ATLAS as representative HEP user communities, we focus on how local data movement operations use hardware- software resources of INFN-CNAF computing centre and which variables affect data movement performances. As a result of this framework analysis we identify GridFTP protocol and GPFS data storage as core services whose performance to study in depen- dancy of typical user defined variables. We therefore decompose data movement commands in operations of increasing complexity i.e., cp, globus-url-copy with and without network, and after defining the mean throughput per file per unit size as target metric, we conduct a quantitative analysis of the contribution and relevance of considered variables across explored usage scenarios.  Finally, we conduct a qualitative fit analysis of the behaviour of chosen throughput metric as a function of relevant user dependent variables. For each scenario and for each variables a best fit model function is selected according to R-square goodness of fit index.