RE: a vote for asymmetric connections in a session

["Y P Cheng" <ycheng@advansys.com>]

From David Robinson
> Sent: Monday, September 11, 2000 11:27 AM
> Having worked on NFS over UDP on high latency lossy links (old ARPAnet)
> the allure of saving a transport level ACK by leveraging the
> fact that the session layer is datagram-like is fatally flawed
> in practice. Works great on a LAN, but when faced with packet
> loss due to congestion the session layer is required to maintain
> adaptive timers, track round trip times, and base retransmissions
> primarily on timeouts. The dropped packet creates huge latency
> bubbles while waiting for timeouts, overly aggressive retransmissions
> further clog congestion, all of these problems are solveable but the
> result looks remarkably like TCP.  In fact this is what we did do to
> NFS over UDP, implementing the various algorithms found in TCP, albeit
> poorly.  Today, most NFS clients speak TCP by default with a minor
> penalty for local servers but a major win for distant servers.

I admit that simply stating UDP will replace TCP in iSCSI was a mistake.
What I tried to point out was the long latency of the ACK's of the TCP
between two endpoints on a WAN can be very costly.  The SCSI protocol by
having the status PDU will ensure "reliable reception."

With or without ACK, a new NIC adapter has taken on a lot of TCP tasks such
as reordering, retransmission, and congestion flow control. (See my other
posting on how the  RTT and data transfer are handled like having an
asymmetric queue.)  In fact, the adapters with VI implementation can easily
perform RDMA for TCP, hence, eliminating the TCP memory-to-memory copy
overhead.  Without a NIC card, we can implement equivalent functions in the
iSCSI driver. I hope in defining the iSCSI protocol we should consider the
power of the new NIC adapters without worrying too much on the old
programming model of TCP queuing and stack processing.