Arvados

Originally https://arvados.org/issues/3699#note-43

I'm copying a pipeline with a 5990M collection. I noticed this code:

                    data = src_keep.get(word)
                    dst_locator = dst_keep.put(data)

See attached image, there's a very clear falloff between blocks -- doing this sequentially isn't optimal. Download and upload could proceed concurrently. Also, it's possible we could get better utilization if we transferred multiple blocks at a time (e.g. 2x down / 2x up) by talking to multiple Keep servers. Consider a producer-consumer pattern using Python queues.