Many real world scientific problems, in fields such as atmospheric science, quantum physics, and structural engineering, are simulated on computers. Due to model complexity, fidelity, or time scales, such simulations often must be run on massively parallel computers. The time and effort involved in designing these simulations is large. %Therefore, many simulations attempt to reuse common operations and methods whenever possible. Libraries containing common operations are typically developed to facilitate simulation design. At Sandia one complex library is Trilinos \cite{IK:Trilinos}. Therefore, many simulations leverage existing optimized kernels and algorithms provided by other software libraries. At Sandia, one such source of state-of-the-art numerical algorithms is the Trilinos project \cite{IK:Trilinos}. Trilinos is a collection of scientific computing libraries called ``packages''. Each package in Trilinos has unique functionality, and is written by domain experts. Packages are typically autonomous, but can leverage capabilities in other Trilinos packages. Functionality available within Trilinos packages includes basic linear algebra operations, preconditioning, solvers, data distribution and load balancing. The Trilinos project provides application developers a suite of modern optimized numerical methods. In turn, Trilinos leverages basic libraries such as the BLAS \cite{IK:BLAS} and LAPACK \cite{IK:LaPACK}. Epetra, a foundational package within Trilinos, is frequently used by other packages \cite{IK:Epetra-site}. Epetra provides fundamental classes and methods for serial and parallel linear algebra. Classes available include point and block matrices, multivectors, and graphs. These and other classes support the usual linear algebra operations. All solver packages within Trilinos can use Epetra kernels as building blocks for both serial and parallel algorithms. For this reason, the performance of solvers depends upon Epetra's performance. Therefore, making Epetra as efficient as possible will improve the performance and efficiency of other packages that depend on it. Just as a program is only as efficient as its underlying components, a parallel program can only be as efficient as the code run on each processor. Even if a program scales efficiently, if its underlying serial code is inefficient, its parallel implementation will be inefficient. By improving the performance of the single-processor portion of a parallel program, the potential top speed of a parallel program is improved. For example, in many scientific programs an important kernel operation is matrix-vector multiplication. By speeding up this kernel, overall simulation speed can be improved. The Optimized Sparse Kernel Interface (OSKI) provides many highly tuned matrix vector multiply kernels \cite{IK:OSKI-paper,IK:OSKI-site,IK:OSKI-user}. OSKI provides five optimized, serial, sparse matrix-vector kernels: four routines that perform matrix-vector multiplication and one that performs a triangular solve of a system of equations. At install time, OSKI's kernels can be tuned according to the underlying machines architecture. At runtime, OSKI's kernels can be tuned according to matrix/vector structure. %By making OSKI's kernels and most of the rest of its interface available in Epetra, we The new Epetra/OSKI interface enables Trilinos and application developers to leverage the highly tuned kernels provided by OSKI in a standardized manner. In this paper, we discuss our implementation of an interface to OSKI within Epetra and assess its performance. In Section \ref{IK:sec:oski}, we give an overview of the design and features of the OSKI package itself. In Section \ref{IK:sec:design}, we discuss the design of the Epetra interface to OSKI. In Section \ref{IK:sec:results}, we discuss the results of performance tests run on the OSKI kernels within Epetra. Tests were run on individual OSKI kernels, and include small scaling studies. In Section \ref{IK:sec:conclu}, conclusions of the work and results described in this paper are presented. In Section \ref{IK:sec:future}, ways to add more functionality to our implementation, and suggestions of things to test in new OSKI releases are presented.