Pulsed Media traced a same-rack slowdown to a netmask too narrow to cover the neighbour

Two hosts have to be on the same wire to talk directly, and being on the same wire does not make them do it. Each host still decides, per destination, whether the address is local or gets handed to the gateway, and the netmask is that decision ^[5]. `ip route get` asks the kernel what it would do with a packet to a given address and prints the answer without emitting a single frame ^[9]. On the hairpinning host it came back as `10.0.0.20 via 10.0.0.1 dev eth0` ^[10]. A working same-segment path prints `10.0.0.20 dev eth0`, and `dev` with no `via` means the kernel will ARP for the destination and deliver it directly ^[11]. The addresses in the write-up are illustrative ^[10].

The gateway carried each byte twice while being configured correctly. Väinämöinen, who describes himself as an autonomous AI sysadmin running in production at the Finnish seedbox host Pulsed Media ^[1], is explicit that a router handed traffic for a destination out another interface is doing precisely what a router does ^[8].

The report puts that one word at 3 MB/s against 55 MB/s ^[12], and the spread across affected paths at roughly 10x to 69x, on the same hardware over the same link ^[14]. 55 divided by 3 is about 18, which sits inside that band ^[23]. In hours, a terabyte between those two hosts takes about 93 hours at 3 MB/s and about five at 55 ^[24]. The before numbers are host-to-host measurements of the broken path, and the after numbers are sustained figures with the direct route in place, measured the same memory-to-memory way ^[15]. The test never touched a disk, so the figures describe the network path only ^[3]. For those multiples to mean anything in another rack, the hosts there have to be hairpinning already, and the read-only lookup settles that before any transfer is timed ^[9].

Stream count is part of the diagnosis. A single TCP flow is loss- and latency-sensitive, which the report attributes to Mathis et al., and a hairpinned, contended path punishes one flow far more than eight ^[16]. "If your box-to-box speed scales with stream count, that is a clue, not a solution," Väinämöinen wrote ^[17].

The on-link route went in live and was measured; a day later the pair was slow again ^[18]. A reboot regression, in the report's reading, means the fix was correct in the running table and wrong where it was persisted ^[19]. The persistence line had been written by a template, which built it with a shell `sed` append whose intended leading tab was expressed as `\t` ^[20]. The published text stops mid-sentence there, and the write-up does not say which character ended up in the file ^[22]. The same one-line lookup, run after a reboot, checks the persisted config instead of the running table ^[9].