It's really bad. The problem is that the swapping of pages is constantly stalling the main Redis thread, which is not built for parallelism. Redis is built around a single highly efficient thread, so stalling it for 250us to pull a page from disk instantly reduces the throughput by an order of magnitude.