Post #4262042

2026-07-11 11:56 UTC

The numbers behind the previous post, for the curious. Same fio and sysbench setup on both boxes, Debian 13, direct=1, libaio. Random read 4K, the workload databases actually live on: 1.41M IOPS on the 2025 box against 822k on the 2027 one. RAM bandwidth: 64,038 MiB/s against 11,906. That is 5.4x. Single thread CPU: 4,504 against 4,147 events/sec. Same EPYC Genoa on paper. And here is my favorite part. The old box ran all of this while carrying my entire production: Mastodon, Postgres, Valkey, a pile of Docker containers, 17 days of uptime. The new box was empty and idle. The loaded veteran still won almost everything. #SelfHosting #VPS #Benchmarks #fio #sysbench #Homelab #MastoAdmin #Fediverse

Replies (2)

@dima@dol.social 2026-07-11 11:59

For the pedants, and I say that with love: the memory numbers differ between the two posts because the test parameters differ. The first run used the sysbench defaults, a 1K block size and a single thread, so it mostly measured per-call overhead. The second used 1M blocks across 4 threads, which measures actual bandwidth. The gap between the boxes holds either way.

Open ##4262043
@dima@dol.social 2026-08-04 12:56

Email from @OVHcloud@mastodon.social this morning: the bandwidth on my VPS went from 400 Mbit to 1 Gbit. Same plan, same price, nothing for me to do. This is the box running dol.social. The one I benchmarked last month, on the grounds that its memory and disk beat the newer 2027 model and the network gap did not matter enough to move. The network gap is now gone too. Correction while I am here: I said 500 Mbit in that thread. It was 400. The conclusion holds, my number did not. #SelfHosting #VPS #OVH #Mastodon #Homelab #Fediverse #Blog #Thoughts

Open ##4375559