SORT BY:

LIST ORDER
THREAD
AUTHOR
SUBJECT


SEARCH

IPS HOME


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

    Re: Symmetric vs Asymmetric



    "VonStamwitz, Paul" wrote:
    
    > > "VonStamwitz, Paul" wrote:
    > > > I agree with David. I oppose a two connection minimum.
    > >
    > [Matt Wakeley]
    > > I disagree. Initial implementations and cheap hardware will
    > > have the iSCSI
    > > implementation in software.  It is easy to do anything in
    > > software - open a
    > > thousand connections, perhaps hundreds of thousands, one per
    > > each LU, it
    > > doesn't care - it just uses up more memory.
    > >
    > > However, for the high performance implementations that will
    > > implement iSCSI in
    > > hardware, doing a function one way sometimes, and another
    > > other times, will
    > > require more hardware and more testing (by both the hardware
    > > vendors and the
    > > value add vendors) to test both functions.
    > >
    > If immediate data on writes is supported, won't you have to support both
    > functions anyways? If the target (or initiator) chooses which connection to
    > transfer data, why can't the command/status connection be one of the
    > choices?
    
    We haven't discussed how immediate write data will work when commands all go
    down one tcp connection.  If commands only go down one connection, you don't
    want them stuck behind (large) data transfers.
    
    > > One connection per LU also has the huge issue of requiring
    > > lots of on-hardware
    > > resources and memory to maintain the hundreds or thousands of
    > > connections that
    > > will be required.... and you want a cost competitive solution?
    > >
    > I stated before that the resource issues preclude mandating a connection per
    > LU. But if resources allow it, it's not a bad way to operate.
    
    The issue is that it's an OPTION.  It is best (and more interoperable) when all
    devices operate the same way.  Leaving this as a option would require both the
    targets and the initiators to figure out how the other guy wants to talk
    (connection per LU or connection per target) and implement, test, etc the
    appropriate hardware/software.
    
    For example, a host driver would have to be written so that it would sometimes
    talk to target connection based targets, and other times to LU connection based
    targets.
    
    >
    >
    > -Paul
    
    -Matt
    
    


Home

Last updated: Tue Sep 04 01:07:27 2001
6315 messages in chronological order