Re: Requirements specification

[julian_satran@il.ibm.com]

David,

The resource consumption I was referring to is not the buffer space but
rather
state maintenance; the common figure I was given is 6k/socket.
Considering also the limited number of sockets in the OS environments
used by today controllers it did not seem a neat solution.
 And as it was pointed out already from this design you can go to one
connection/LU
while the reverse does not hold.

There is also a structural advantage of connecting to a target not a LU -
the LU map
may change while you are working.

If you leave all the multiplexing to TCP - including using several physical
connections
I am not aware of any good technology to do it today (as TCP traffic is
bound to the
end-point IP addresses). We could decide to throw back the balancing need
to the TCP
implementers but I am not sure that this will get us very far.

Availability is not an issue for parallel SCSI and it is not there (there
are no recovery mechanisms on SCSI) ;  FCP designer found out about this
quite late in the game and moved to FCP-2 (that introduces a fairly complex
recovery mechanism).

Julo

David Robinson <David.Robinson@EBay.Sun.COM> on 04/08/2000 21:37:51

Please respond to David Robinson <David.Robinson@EBay.Sun.COM>

To:   ips@ece.cmu.edu
cc:    (bcc: Julian Satran/Haifa/IBM)
Subject:  Re: Requirements specification




> The one additional requirement is availability/fault-tolerance.

I agree that we need to establish the availability requirements
of IPS, but I see them as orthogonal to the connection discussion.
IP already has the wonderful rerouting properties to be fault
resilient as well as link level technology to do failover.  From
an IPS perspective the host adapters and the storage adapters are
the weak links.  This to do a reasonable HA solution you will need
multiple adapters on both ends which will imply initiator/target
pairs. So I don't see how having multiple TCP connections per session
will increase availability. Existing HA solutions that layer on top
of FC or parallel SCSI could just as easily layer on top of IPS.

> Your arguments about performance are valid. However I doubt that there
will
> be enough incentives - beyond price - to develop things for high end
> controllers and
> servers.
>
> Enabling multiple connections brings those applications the performance
> required
> without any serious implications to the rest of the "family" (as I
outlined
> in Pittsburgh
> controllers and servers that don't need multiple connections/session
don't
> have to implement them).

I am not sure that I understand your point here. As has been mentioned by
others, one performance concern is how to maximize utilization of the
available link bandwidth with a higher level protocol, fundementally
you can never get 100%. Part of the concern is that most existing TCP
implementations are not scaling up as fast as the link layer protocols.
To handle this situation people consider some form of aggregation
of TCP links to relieve the pressure on the implementation side.  The
question becomes what is the appropriate way to aggregate. If we run
multiple LUNs over the same session then we will push the TCP
implementation
to its limits.  However, if we run a session per LUN, the bandwidth
pressures will be dramatically less on TCP, we will more easily
utilize existing and future link layer bandwidth, and simplify the
multiplexing by leaving it in the TCP layer.

> Storage traffic requirements will always exceed those of many other
> applications.

Storage will always exceed thoses of *some* other applications, it is
easy to find numerous applications that exceed storage requirements.
I am not clear on the point, but we need to make sure that we clearly
define what the storage traffic requirements are.

> As for the "one-connection-per-LU" we covered this solution in long
> discussions
> and even several full fledged implementation - as it is compelingly
simple.
> However the resource consumption is unjustifiably high and the security
> problems are
> even worse (the LUs "viewed" by an initiator depend on who he says he is)
> than
> in the current draft.

I don't agree that the security problems are worse.  For any session,
whether
it is a session per target or a session per LUN will need to do full
security negotiation.  Realitive to the expected lifetime of the sessions
the time and complexity will be negligible. If we adequately solve it for
one model it should trivially apply to the other.

I also don't agree on the resource consumption being too high. In terms
of memory resources the necessary data buffering is dependant on the
aggregate flow between the target and initiator which will be
independent of how many connections were used to get the data to
the other side. There will be more connections, but within typical
TCP implementations the size and complexity of control blocks is
minimal.  The ability to switch and manage many connections rapidly
is a long solved problem. It would be illumninating to hear if Adaptec
has found that their connection per LUN is resource consuming.

     -David