SORT BY:

LIST ORDER
THREAD
AUTHOR
SUBJECT


SEARCH

IPS HOME


    [Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

    Re: ISCSI: Urgent pointer consensus



    Matt Wakeley wrote:
    
    > Daniel Smith wrote:
    >
    > > Matt Wakeley wrote:
    > > >
    > > > placed.  If you lose an iSCSI PDU header due to a lost TCP segment, you
    > > > lose iSCSI PDU framing from then on (until the missing segment is
    > > > received).  You then have to store the TCP data (received after the
    > >
    > > Ah!  Now I am beginning to see how your implementation works.  In your view,
    > > a write CDB for 128MB would have a sequence like...
    > >
    > > send CDB PDU
    > > rx R2T for 1MB          (rx=receive)
    > > send 1MB
    > > rx R2T for 1MB
    > > send 1MB
    > > ...
    > > rx status PDU
    >
    > Actually, I see something more like:
    >
    > send CDB PDU A
    > send CDB PDU B
    > rx r2t for A
    > send CDB PDU C
    > rx data for B (512-4K typically)
    > send data for A
    > rx data for C
    >
    > In other words, lots of I/Os with lots of small (no longer than a few K) iSCSI
    > messages going in both directions.  Typical disk I/Os are not in the megabytes.
    > That way, if "rx data for A" gets lost, the data for B and C (and all the others
    
    This should say "rx r2t for A" gets lost...
    
    >
    
    >
    > that will arrive - there could be 10's or 100's of commands outstanding) can be
    > placed in their dedicated I/O buffers before A is retransmitted.
    >
    > -Matt
    >
    > >
    > >
    > > ... whereas some of us imagine something like...
    > >
    > > send CDB PDU
    > > rx R2T for 128MB
    > > send 128MB
    > > rx status PDU
    > >
    > > For short transactions (<1MB) the urgent doesn't help much because R2Ts will
    > > cover the whole amount of data.
    > >
    > > For long transactions (>1MB) there may be some benefit, although with all
    > > those R2Ts flying about TCP may lose streaming metrics.  Hmm, I'll give it
    > > more thought, but this still seems very specialized and closely tied in with
    > > your (and others) implementation view.
    > >
    > > Daniel Smith.
    > > --
    > > IBM Almaden Research Center, 650 Harry Road, San Jose, CA 95120-6099, USA
    > > K65B/C2 Phone: +1(408)927-2072 Fax: +1(408)927-3010
    
    


Home

Last updated: Tue Sep 04 01:06:25 2001
6315 messages in chronological order