RE: LU access through an iSCSI session

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

Matt,

As you are clearly designing an interface for an on-site controller, why use
a large numbers of TCP connections?  For low latency, would an ST like
protocol be a better choice?  See http://www.ietf.org/rfc/rfc1819.txt.
Confirming placement of the controller being at the client, you have little
concern about the treatment of the network which is obviously dedicated.
Why bother with IETF?  There are many alternative interconnects and more
coming, why re-invent this interface if it is indeed local and dedicated?
Surely, even with hardware assist, you will be better off with other
solutions.  My idea of an ideal controller would be one that bolts into the
system and finds access via SCTP.  Intel will suggest Infiniband.  What does
Sun, IBM and HP use to talk to local controllers?  Fibre-Channel?

Doug

> -----Original Message-----
> From: owner-ips@ece.cmu.edu [mailto:owner-ips@ece.cmu.edu]On Behalf Of
> Matt Wakeley
> Sent: Wednesday, September 20, 2000 10:09 AM
> To: IPS Reflector
> Subject: Re: LU access through an iSCSI session
>
>
> The talk of making "lots of TCP connections look like a big pipe"
> is NOT meant for
> congestion control.  It is meant to obtain higher bandwidth on a
> local network, such
> as in a data room.
>
> For example, one could have multiple 1Gbit TCP connections from a
> host to say a
> switch.  The switch might be connected via a very high speed
> backbone to another
> area where the storage devices are, where there is another such
> switch.  Similarly,
> a disc array may be connected via multiple 1Gbit TCP connections
> to that switch.
>
> There is lots of early discussions on why multiple TCP
> connections are required to
> obtain parallel paths through the fabric.  You might check out
> the email archives,
> especially messages posted by Randy Haagens.
>
> -Matt
>
> Randall Stewart wrote:
>
> > Matt Wakeley wrote:
> >
> > > .....snip....snip....
> > > Again, the *only* answer is "Yes".  The multiple TCP
> connections make up a
> > > *single* session.  The session just happens to have lots of
> "TCP connections"
> > > to make it look like it has a "big pipe" (for example, 4
> 1Gbit connections can
> > > be used to emulate a 4Gbit link).
> > >
> > > If you want different "LU views" per TCP connection, then
> each TCP connection
> > > would be a separate, independent iscsi session.
> >
> > ....snip....snip.....
> >
> > The above talk of making "lots of TCP connections look like a
> big pipe" sounds
> > once
> > again like attempting to get around congestion control in the
> network... I know
> > we
> > had this conversation before.. but what is stated above
> strongly implies this
> > .... and
> > I don't think this is the behavior one would want on the Big
> I-internet. I can
> > see
> > a private network doing this ... but I don't see how it could
> be supported in a
> > standard manner that can be used on the Iternet...
> >
> > R
>