Re: iSCSI Markers

[Mark Bakke <mbakke@cisco.com>]

Glenn-

I agree that TUF is a better solution for iSCSI, and do
not believe that either TUF or Markers are needed for this
first version of the iSCSI protocol.

I believe that the best solution so far for iSCSI framing is
in TUF/PDU Alignment or to use another message protocol instead
of TCP.  Adding markers using either method is not usable by
any other application than iSCSI, and adds some demands on
implementations that can not make use of it anyway.

Regardless of whether it is standard, TUF is not that hard
to do, and it's what we should be doing.  Without a more
general solution like TUF, TCP will start running out of gas
for other applications as well as iSCSI, and we will have to
go look at other protocols instead when we get to 10G and
beyond.  I don't think that modifications to the usage of
TCP are outside the scope of what iSCSI should influence.  We
clearly can't just do the work ourselves in a bubble, since
that's not our charter.  But we can certainly influence other
high-bandwidth TCP users to help come up with good, general
solutions.  I don't think that the cost/performance problems
that markers are trying to solve really exist today.  We can
deal with the performance issues at the expense of more
memory on NIC cards, or by doing scatter-gather on storage
devices and gateways.  It just costs a little more.  As the
market for iSCSI stuff develops, and costs have to go down,
we can address this by doing the right solution, which I
do not believe involved markers.  iSCSI is not the only
protocol that is used to carry data on a network.  NFS, CIFS,
DAFS, FTP, HTTP, and others carry far more data than iSCSI.
If we don't solve the problems for these as well, our NIC
offload cards will eventually have to have "special" code
for each one of these, which may end up being just as
expensive than adding some memory.

I personally think that there is no reason to mandate markers
or framing of any kind in this iSCSI RFC.  We are just not
ready, and we already have enough things (security) that are
mandated that will not be part of most implementations anyway.
We are just not ready for the real solution yet.

Those who wish to use markers for the time being should
certainly do so, but we must not (oops, should have said
that in caps :-) mandate their implementation and use.
If markers stay in the draft in either form, they should
remain optional.

Just my opinions,

Mark

Glenn Dasmalchi wrote:
> 
> > -----Original Message-----
> > From: John Hufferd [mailto:hufferd@us.ibm.com]
> > Sent: Thursday, January 10, 2002 1:27 AM
> > To: ips@ece.cmu.edu
> > Subject: iSCSI Markers
> >
> > I see no relationship to COWS, except both have an 8 byte
> > header.   If you
> > have TUFs, you do not need to have Word Stuffing.  The
> > problem with COWS is
> > not is header it is the Word Stuffing and the Replacement Pointers.
> >
> The random key marker approach in the TUF draft has a non-zero
> probability of generating a "false positive" in theory (requires
> resegmenting middle boxes + aliased data + exact length resegmentation,
> etc.). Personally I think the probability is so low as to not be a
> consideration, but a COWS approach using TUF-style segmentation does
> address the issue.
> 
> The point about the iSCSI charter is interesting, do others feel that
> the segmentation requirements for TUF/PDU Alignment disqualify it from
> consideration for iSCSI?
> 
> -Glenn

-- 
Mark A. Bakke
Cisco Systems
mbakke@cisco.com
763.398.1054