Re: "Lower overhead" iSCSI

["Julian Satran" <Julian_Satran@il.ibm.com>]

Dan,

We don't want start this discussion - it is out of the charter of this group.
A bit of advise - before even getting to the public - we at IBM did run a testbed with the following results:

• no advantage for UDP on any relevant performance metric (including CPU utilization) even before considering the effects or recovery code
• on most stacks it performed worse (rumor has it that stacks have TCP very well tuned, not so UDP)
  
  Adding to that the lack of congestion control took it out of our vocabulary.
  
  DCP if successful and widely deployed may change that in the future. Or SCTP may.
  
  Julo
  
  
  
  
  

  Dan_McConnell@Dell.com Sent by: owner-ips@ece.cmu.edu 30-01-02 00:29

  To:        ips@ece.cmu.edu         cc:                 Subject:        "Lower overhead" iSCSI

  
  
  Let me start off by saying that I am interested in doing "iSCSI" protocol
  over UDP.  Now I realize that this is an old issue and will probably start
  some "religious" battles, but let me state the scenarios before I receive
  death threats.  The planned environment that this will go into is a small
  one with say 10 servers connected through a "non-blocking" switch to the
  storage device (ie no routers, gateways, etc...just direct point-to-point
  connections).  This is assuming that the switch is really non-blocking and
  hopefully implements flow control or pause frames.  So technically all you
  should have to worry about is port/device contention.  However, when you
  think about it...this is similar to FC.  FCP runs on class 3 FC which is a
  non-reliable transport protocol such as UDP and handles contention, also
  some of the early "SAN interconnect" guys are doing this today with
  relatively good performance and few issues.
                  The attempt here is to maintain low CPU utilization at high
  performance rates.  While I realize that these TOE devices are moving along
  rapidly, there are some situations where they are not feasible, such as a
  blade server environment (no PCI slots, and no real estate/power available
  for onboard TOE).  Worst case scenario is that a packet is dropped or
  received out of order and the ULP (SCSI) must resend the cmd/data sequence -
  still no data lost, just a temporary performance hit.
                  So my question is:  is this feasible?, and why not implement an
  "iSCSI" protocol layer that can run over TCP or UDP(though I realize it
  won't be considered "standards compliant")?
  
  Thanks,
  Dan