System performance modelling WG meeting
- Local: Andrea Sciabà, Evangelia Santorinaiou, Domenico Giordano
- Remote: Renaud Vernet, Catherine Biscarat, Gareth Roy, Oxana Smirnova, Pepe Flix, Andrea Sartirana
Renaud presents a summary of the information we got from some Tier-1 centres in reply to his questionnaire on site infrastructure costs (refer to the slides). About disk costs, he pointed out that comparing sites is not straightforward because they might use different technologies and hardware with different performance. For tapes, a site reports a much higher cost per TB of media but it was the price from 2016, while the others report the 2018 prices.
Little information was given about cost evolution; in case nothing else is received, the cost evolution might be taken from the estimates of the new HEPiX technology watch working group, or of Bernd Panzer-Steindel.
Pepe asks if sites were requested to provide estimates for costs they do not directly pay for. The answer is no, "indirect" costs are not estimated. Pepe mentions that in his study for NDGF a federated model turned out to be preferrable exactly because many costs are born by universities, who act as sorts of "sponsors" (and thus, better to have more!). On the other hand, Renaud argues that just ignoring e.g. electricity costs because they are pad by a university does not mean that they don't exist, because if the site expands and the electricity bill increases, the university might now be willing to pay it. Oxana adds that this very situation happened for real.
Pepe remarks the need to add manpower costs, but Renaud replies that not considering them was a deliberate choice.
Andrea Sc. adds that it is important to take note of all the differences from site to site (including some costs being "hidden" at some sites), but at the end, to have a workable model, we will need a simplified view of the costs, not too detailed but realistic enough to be used for evaluating different scenarios.
It was decided to pass the questionaire to Gareth and Andrea Sa. to give their opinion, in view of a possible extension of the survey to selected Tier-2 sites (certainly not to all WLCG sites).
Domenico gives a talk on the strategy of the HEPiX benchmarking WG for packaging experiment workloads, based on a new tool soon to be available in CVMFS. It was successfully applied to KV and the CMS ttbar simulation and it will be soon applied to some of our reference workloads.
During the task list review, Andrea Sc. reported that Concezio has prepared a new LHCb reference workload and Andrea successfully tested it; however, it does a lot of event filtering at the generator level, which is not relevant for our needs, and he will soon provide a version that does not do any filtering.
Andrea Sc. and Renaud reminded the audience that next week a meeting of the subgroup on cost estimation has been called, to review in more detail the answers to the questionnaire. People interested in participating are invited to read before the meeting the answers that Renaud distributed in the past days.