Re: iSCSI: Markers

[Paul Koning <ni1d@arrl.net>]

>>>>> "John" == John Hufferd <hufferd@us.ibm.com> writes:

 John> Paul, I did not understand your statement ...

 John> "...It seems to me a node needs at least as much memory in the
 John> other direction -- quite possibly more if it's the storage end,
 John> given that a lot of the traffic is reads.)"

 John> Are you talking about all the memory in the Node or the RAM on
 John> the HBA?  The focus should be on the HBA, and generally, that
 John> memory is small in the outgoing direction.  So perhaps I
 John> misunderstood your point.

I meant the HBA.  Would you expect TCP level retransmit to fetch the
data from the host again?  Yes, I guess then you could save the memory
on the HBA.

 John> About the Mandating of markers.  The proposal was that it
 John> should be required to implement (not required to use) in
 John> outgoing, and clearly it should be optional to use.  In that
 John> manner, the approprate trade offs can be made at execution
 John> time.

Yes, I know.  I'm objecting to the "required to implement".  It's
problematic to have a feature that someone else can turn on (by asking
for it when you're required to say "yes") and which has the side
effect of destroying your performance.

 John> It seems to be a small impact to the code to just to support
 John> this on the outgoing flow.  I did understand you comments about
 John> how some HW operates, and that needs to be considered in the
 John> executions time trade off.  In most Desktops, laptops, that is
 John> probably not an important issue (with 1.5-2 GHz processors),
 John> but it could be a serious consideration for any servers that
 John> need to have a SW implementation.  (Of course, many folks would
 John> tell the servers to get an HBA, but that is a different issue.)

I'm talking about embedded systems, where throwing an outboard HBA
into the picture makes things much bigger and more expensive, not to
mention slower.  In that example, turning on markers changes the
transmit flow from a straightforward DMA of the data into a memory to
memory copy with insertion of the markers every N bytes.  That's a
large increment in overhead.

      paul