The diminishing returns from Moore’s law and technology scaling have significantly driven the deployment of a plethora of accelerators in large scale computing infrastructures. While the design of such accelerators is being broadly addressed, the challenge of designing an intelligent scheduler that achieves optimizes performance while abstracting low-level systems details for the heterogeneous processing fabric from the programmer continues to be a major problem. An important application where the foregoing challenges come together is computational genomics, particularly the “Variant-Calling and Genotyping Analysis.” The Symphony system designed and implemented at Illinois (and the object of this paper) reduces this challenge to that of utilizing Bayesian inference in probabilistic graphical models to tie together heterogeneous compute resources like general purpose accelerators (e.g., GPUs, MICs), and custom designed application specific processors (prototyped on FPGAs). We demonstrates the performance, resource-management, scaling, scheduling and isolation properties of Symphony. Our evaluation shows that the use of Symphony improves overall benchmark performance (on a single node) from 73 hours to under 45 minutes (88x and nearly 210x in terms of performance-per-watt) for human genome datasets that are appropriate for clinical use.