From RFC 4253:
Each packet is in the following format: uint32 packet_length byte padding_length byte[n1] payload; n1 = packet_length - padding_length - 1 byte[n2] random padding; n2 = padding_length byte[m] mac (Message Authentication Code - MAC); m = mac_length [...] random padding Arbitrary-length padding, such that the total length of (packet_length || padding_length || payload || random padding) is a multiple of the cipher block size or 8, whichever is larger. There MUST be at least four bytes of padding. The padding SHOULD consist of random bytes. The maximum amount of padding is 255 bytes.
Why does SSH require (or recommend with SHOULD) random padding, as opposed to non-random padding?
And why does RFC 4344 say that it is not necessary when using CTR mode?
As an additional note, when one of the stateful-decryption counter mode encryption methods (Section 4) is used, then the padding included in an SSH packet (Section 4 of [RFC4253]) need not be (but can still be) random. This eliminates the need to generate cryptographically secure pseudorandom bytes for each packet.