RE: iSCSI: Flow Control

["Douglas Otis" <dotis@sanlight.net>]

Matt,

The present iSCSI draft assumes each transport may perform a proxy or
redirection (via either IP or SCSI media) at login of the client.  In the
case of redirection not over IP, SCSI medium is not considered unless each
target is to act in the role of Service Delivery Port, and LU takes the role
of target where LUN is fixed.  This virtual target for a medium is not
possible unless everything on the medium is restricted to the same device
type.  In the case of redirection over IP, the proxy will authenticate
locally and on the forwarded target over yet another connection should this
proxy-target connection conform to the proposal.  What could have been a
single connection using an encapsulated target/initiator descriptor becomes
tens or thousands of separate flows to be performed.  In addition to being
expensive overhead, it is not network friendly when each connection or
connection pair is used to resolve a target.  As each target can implement a
proxy service, a path will be difficult to define for either IP or SRA due
to a desire to obfuscate actual underlying transport.  Despite standard
encapsulation schemes for multiple connections, the desire to add additional
connections is sure to make this unfriendly transport even more aggressive.
All of this still overlooks flow control as required, at the medium.
Perhaps this architecture is intended for Infiniband.  Without target
resolution, resources are consumed by the slowest hidden device.

Doug