A descent algorithm is given for solving a large convex program obtained by augmenting the objective of a linear program with a (possibly nondifferentiable) convex function depending on relatively few variables. Such problems often arise in practice as deterministic equivalents of stochastic programming problem. The algorithm s search direction finding subproblems can be solved efficiently by the existing software for large-scale smooth optimization. The algorithm is both readily implementable and globally convergent.