1: \begin{abstract}

2: We describe an asynchronous parallel stochastic coordinate descent

3: algorithm for minimizing smooth unconstrained or separably constrained

4: functions. The method achieves a linear convergence rate on functions

5: that satisfy an essential strong convexity property and a sublinear

6: rate ($1/K$) on general convex functions. Near-linear speedup on a

7: multicore system can be expected if the number of processors is

8: $O(n^{1/2})$ in unconstrained optimization and $O(n^{1/4})$ in the

9: separable-constrained case, where $n$ is the number of variables. We

10: describe results from implementation on 40-core processors.

11: \end{abstract}

12: