My first suspicion is the multiprocessing capability while using numpy library may be redundant and can cause slow down. If that makes sense to you, you can restrict the multiprocessing for each Numpy instance. However, if you think the bottleneck is at numpy file reading operation, then you may like to see if a scanner is slowing the operation down or perhaps the lcoation of the production is not conducive for fast operations. You can isolate the read operation and time it in production to rule it out.