Question about AllReduceEA

[sidps]

From the code and the algorithm presented (https://github.com/twitter/torch-distlearn/blob/master/lua/AllReduceEA.md), it seems like the all-reduce step involves synchronization between workers.

The algorithm published in the original paper does not require such synchronization (sec 3.1: each worker maintains it's own clock, t_i). Is AllReduceEA then an algorithm for synchronous EASGD, for which only the formulation was presented in the paper (sec 3)?

If so, are there any comparisons between synchronous and asynchronous EASGD?

Apologies if I have misunderstood this.