Re: iSCSI: Urgent Pointer Negotation

[Matt Wakeley <matt_wakeley@agilent.com>]

John Hufferd/San Jose/IBM wrote:

> Please forgive my previous send of a null message (had a finger check).
>
> Costa MIGHT have the key to this discussion.

Well, I already suggested negotiation of the urgent pointer (see ISCSI:
negotiation of the use of the Urgent Pointer), but I think Costa phrased it
better than I did.

>  If he is correct, that the
> Urgent Pointer takes the fast path away from the SW receive TCP Stacks, but
> perhaps not a big deal on Send (Need to check this),

As Mark said, it will require two calls to send(), but perhaps that isn't as
big a deal as the inbound path not taking the fast path.

Mark said:

> Maybe I am being simplistic, but I look at the impact that this has on the
> sending side in order achieve this.  It means that for every iSCSI PDU I
> have to call send twice, that is:
>
> send_iSCSI (char *  buf,  int    len)
> {
>    send (buf, 1, MSG_OOB);
>    send (buf + 1, len - 1, 0);
> }
>
-Matt


> then I would think
> that SW implementations could almost always agree on send, but not on
> receive.  In that way, a HW TCP/IP with iSCSI NIC could get just about all
> it needed in performance improvement and Memory reductions without a
> significant impact on the SW side. Therefore, this MIGHT be a break
> through.  We need to confirm his statements, and then,  it might make since
> to have, as Costa suggested, a Login negotiation parameter for Urgent
> Pointer, in each direction.
>
> Those that really know, about the receive fast path and the send  path with
> Urgent Pointer stuff, please answer quickly.
>
> .
> .
> .
> John L. Hufferd
> Senior Technical Staff Member (STSM)
> IBM/SSG San Jose Ca
> (408) 256-0403, Tie: 276-0403
> Internet address: hufferd@us.ibm.com
>
> csapuntz@cisco.com@ece.cmu.edu on 11/15/2000 06:33:16 PM
>
> Sent by:  owner-ips@ece.cmu.edu
>
> To:   ips@ece.cmu.edu
> cc:   csapuntz@cisco.com
> Subject:  iSCSI: Urgent Pointer Negotation
>
> If the URG pointer feature is to be used, it should be negotiated.
>
> The reason it should be negotiated is that the presence of the URG
> pointer takes many TCP receive stacks off of their fast path. I am
> not familiar with whether it significantly slows current TCP send stacks.
>
> Each half of the connection should be negotiated separately. This
> deals with a hardware accelerator on one side and a software
> implementation on the other.
>
> Cheers,
> -Costa