Genetic recombination can speed up or slow down evolution toward the global optimum. For analyzing mutational trajectories to the optimum it is conventional to study peaks in the fitness landscapes. However, such mutational peaks cannot provide a complete picture of the interplay between mutations, recombination and selection. Here we use recombination peaks as an additional tool. Informally, a recombination peak is a set of genotypes such that no allele shuffling among members can improve average fitness. If conventional peaks are optimal with respect to mutations, then recombination peaks are optimal with respect to recombination. The approach applies to general fitness landscapes with an arbitrary number of alleles and loci, and our experiments involve up to five alleles. One of the most frequently observed patterns of epistasis in nature is diminishing returns epistasis, i.e., the combined effect of beneficial mutations is smaller than a multiplicative expectation would predict. Recombination tends to speed up adaptation toward a peak whenever there is a systematic pattern of diminishing returns epistasis near the peak. The opposite is true for increasing returns epistasis. Mixed curvature is more challenging to analyze. Recombination peaks provide a language for describing effects of recombination, and threshold results are discussed accordingly.