On 5/15/12 5:20 PM, Galder Zamarreño wrote: > On May 15, 2012, at 4:58 PM, Paolo Romano wrote: > >> Hi Galder, >> >> Let me try to clarify this. >> With Diego we have developed a system for >> forecasting the performance (e.g. maximum throughput, abort rate, avg. >> transaction execution time) of an ISPN application when it is deployed >> on a cluster of a different scale (compared to the current one). >> >> We modelled using analytical techniques and machine learning the locking >> and network-related behaviors of ISPN, but we did our work on ISPN 5.0 >> (replication mode), which used this "hybrid", partially local/partially >> remote (and distributed=> no primary owner). >> >> In other words, our performance prediction schemes won't work neither >> for Optimistic nor for Pessimistic. > Why do the predictions don't work? Because the locking scheme has changed, hence its performance are changed, hence we are forecasting something different now :-) > > Where you guys somehow overriding some of the existing logic in 5.0 to track > or record some data? We were collecting additional statistics on tx usage. We have discussed this with Mircea, who was in our last meeting in Lisbon with us, and he has suggested how to re-implement that part to minimize intrusiveness of our software probes.
Indeed, as we were discussing with Mircea and Sanne, we would be glad to contribute these statistics to Infinispan as they may be useful not only for our performance forecasting techniques, but also for human-driven performance tuning. We are re-testing this code right now, and we will have a stable version at the end of the week. _______________________________________________ infinispan-dev mailing list [email protected] https://lists.jboss.org/mailman/listinfo/infinispan-dev
