sftp: speed / latency / protocol issues

[ThomasWaldmann]

Documenting some fundamental issues of the sftp protocol:

• sftp protocol uses 32KB blocks for data transfer and usually each block has to get acknowledged individually, which slows down transfers significantly on connections with notable latency.
• paramiko offers pipelining (write) and prefetching (read) to optimize this and borgstore uses these features.
• but in the end, sftp is just not a fast protocol and paramiko is not a fast sftp client implementation either.

Thus, using sftp is not recommended for big stores nor for high latency connections.

We can't do much about that in borgstore besides offering alternatives that use other protocols.

[ThomasWaldmann]

Interesting: https://rclone.org/sftp/#sftp-chunk-size

max_packet_size there:

https://docs.paramiko.org/en/latest/api/transport.html#paramiko.transport.Transport

[ThomasWaldmann]

From paramiko/paramiko#2438 (comment) :

sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
sock.setsockopt(socket.IPPROTO_TCP, socket.TCP_NODELAY, 1)

sock.setsockopt(socket.SOL_SOCKET, socket.SO_SNDBUF, 32 * 1024 * 1024)
sock.setsockopt(socket.SOL_SOCKET, socket.SO_RCVBUF, 32 * 1024 * 1024)

sock.connect((self.host, 22))
self._transport = paramiko.Transport(sock=sock)