Den ons 20 maj 2020 kl 12:00 skrev Ignazio Cassano < ignaziocassano(a)gmail.com&gt;gt;: Perhaps most of the time is lost talking over the network, making the small differences in how well ssd -vs- nvme perform not visible. -- May the most significant bit of your life be positive.

Den ons 20 maj 2020 kl 12:14 skrev Ignazio Cassano < ignaziocassano(a)gmail.com&gt;gt;: I think there are several factors to weigh in, when you need to maximize performance, from putting BIOS into performance mode, having as fast network as possible (some say 40 is 4x10 so 25Gbit is still faster in latency than 40GE, but don't quote me on that) and so on. Also, in my experience, nvmes can eat lots more parallel writes than ssds, so a single test might just be easy enough for both drives to accept, whereas if you had 100s of writers at the same time, it might favor the nvmes who would not slow down as fast as ssds would in such a case. Perhaps the nvmes will be far better everytime the rocksdb does level-compactions so that it doesn't give you better write numbers in tests, but less stalls when/if the DB needs reorganizing. On the other hand, if writes always are huge and evil and long-lasting, sooner or later you will be capped by how much data the backend disks can take, regardless of the performance of nvme/ssd WAL/DBs. Doing both local and remote (which you did) benchmarks, looking at iostat for both nvme and ssds while doing the test and perhaps most important of all, knowing when an acceptable result is achieved. If you have a target to reach and you reach it then perhaps moving on to other parts of designing is best. -- May the most significant bit of your life be positive.