I have been trying to reproduce it with a localhost samba and I wasn't able to. I also tried adding delays and limiting bandwidth using tc with the idea of replicating network delays of a gigabit ethernet network but I didn't see a significant difference either. So I investigate what is the difference between writing to the NAS (where I discovered the problem) compared to the samba at localhost and I see a completely different behaviour in the amount of data carried by each SMB message. writing to NAS ----------------------------- read/write 64kB 64kB read/write 1MB 1024kB splice 64kB 4096B splice 64kB (PATCH) 64kB splice 1MB (PATCH) 1024kB writing to localhost samba --------------------------- read/write 64kB 1MB splice 64kB 1MB splice 64kB (PATCH) 1MB In every case, the server seems to send one ack SMB message for each block of data and client doesn't sends new data until the ack is received. I suppose this explains why the throughput is ridiculous low for the situations when a SMB message carries so little data.
