What techniques are NOT used by Petuum to support both data$-$parallel and model$-$parallel machine learning algorithms? $($A$)$ Bounded asynchronous consistency model helps to ensure data$-$parallel convergence guarantees without explicit network synchronization $($B$)$ Global model state can be accessed via distributed shared$-$memory interface $($C$)$ Workers have to access the data in the batch order $($D$)$ Scheduler enables user$-$ fined application consistency rules, thus permitting fine$-$grained control over parallel ordering of model$-$parallel updates.