In your book "Building Storage Networks," page 358 line 12 from the bottom: "Class 2 [ Fibre Channel] is a connectionless service that provides multiplexed communications with acknowledgment of transmitted frames," Page 543, last two lines "Unreliable connections do not acknowledge transmission and are similar to Fibre Channel's class 2."
As the acknowledgement is the main differentiator between class 2 and class 3 Fibre Channels, the lines above seem to be in conflict.
It took me some time to find the reference you cited as page 543. It was actually on page 452. It would have helped if you had referenced IB in your question, but it was sort of fun working through my own maze again.
In the wiring chapters of the book I tried to make comparisons between the various technologies to help readers understand the similarities and differences of the technologies. Unfortunately, this turns out to be pretty difficult when working with IB and FC. (It turns out to be difficult when comparing IB to anything else for that matter.) Perhaps it was a mistake to try, but I wanted to provide a bit of a cross-reference context. So I'm sorry if this was misleading - that was not the intent. Frankly, I admit that IB has been a confusing topic for me, partly because no implementations exist yet and the only sources are the draft specifications that have been published. These are not always the easiest documents to interpret. .
"Unreliable" in InfiniBand refers to the way error recovery is handled. Specifically, in IB, "unreliable" means the network does not provide error recovery. This is similar to both FC class 2 and Class 3 - neither of which provides error recovery in the network - instead error recovery is handled by higher-level functions such as network application drivers. (Error recovery in FC is typically performed by SCSI device drivers after time out thresholds are exceeded.)
IB can have both connection and connectionless unreliable service. In other words, connection-ness is not explicitly tied to error recovery, as my comments in the book seem to suggest. Someday I would like to get three or four networking engineers (1 FC, 1 IB, 1 GbE, 1 ATM) in a room together to discuss this concept just to see if there is agreement among them on whether or not wire level error recovery is a prerequisite for connection-oriented services. I'm quite sure it would not be a slam-dunk discussion.
So anyway, thanks for finding this and bringing it to my attention. It's another example of lurking ambiguity at the bottom of difficult stuff. It makes all our work more difficult, but in the end more interesting. Again, I apologize if this has caused you difficulty in your work.
Also, if you are interested in InfiniBand education, I know of a company called the I/O Authority that is now offering classes.
Editor's note: Do you agree with this expert's response? If you have more to share, post it in our Storage Networking discussion forum at http://searchstorage.discussions.techtarget.com/WebX?replyToMessage@7.uALMagHuauI^1@.ee83ce4!viewtype=convdate
This was first published in November 2001