LMDB is a really great read optimized key-value, transactional key-value store, but you have to be careful with when benchmarking memory mapped files. The problem is that a thread touching paged out memory just blocks without performing any system call. It looks like you're faster and it starts out very fast because as long as your data fits into main memory you can read it without any context switches, but as soon as you get under even slight memory pressure jitter goes through the roof and your userspace code is helpless because there is no load-if-mapped instruction to avoid blocking the thread if your accessed page has to be read in from the backing storage. Your database has no control over I/O parallelism and queue depth except very indirectly by delegating database accesses to a thread pool of a specific size and hoping for the best.

is it only me or does this site works only over plaintext http, while you gave an https link?