some suggestion for improving PS performance

[tanglizhe1105]

when LDA training iteration going, the term-topic array gradually become sparse array, there will appear many zero elems.
So, i wanna to know is there some plans to change model store as sparse array or sparse matrix, if use sparse matrix store, the will definitely reduce communication cost as we use sparse ways fetch parameters & push updates.
As me to know, Microsoft's LightLDA also use sparse way to store and transit model data.

Thank you