If you use grep in OS X, and you use it a lot (even called through other programs/libraries) it pays in time to install GNU grep vs. the default OS X provided one. It's several times faster. Ran into this when benchmarking grep + GNU Parallel on OS X vs. Linux in my thesis (raftlib modeling) work. Results on my small MacPro cluster didn't match those on Linux. The OS X runs were orders of magnitude slower despite almost identical hardware configurations. The culprit as it turned out was largely the implementation of grep. I compiled with the GNU version, re-ran and the rest of the mis-match was simply due to the lack of POSIX features like thread pinning, and NUMA control.

<stepping up to soap box/> So yeah...as usual, algorithm matters. Implementation matters, and well....knowledge of the architecture/data-movement matter. These days the architecture/data-movement knowledge matters almost as much or more than the algorithm. Picking an algorithmically faster implementation that moves data more (e.g., tree traversals usually have poor cache behavior, more bursts required, etc.) is often the case b/c programmers don't really understand how the computer works, but are great at coding O(n logn) algorithms. We shouldn't stop teaching algorithms, but data-movement/access likely needs to be emphasized along with the understanding of algorithms.