SORT BY:

LIST ORDER
THREAD
AUTHOR
SUBJECT


SEARCH

IPS HOME


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

    Re: SCSI URL scheme [WAS: Re: iSCSI: 2.2.6. Naming & mapping]



    Robert Snively wrote:
    > 
    > Textual names are practically useless as identifiers, as witness the
    > challenge of finding "Bob Smith" in an American phonebook.  Every
    
    This is not quite the same problem.
    
    1) You don't get a choice as to what you call your URL ("Robert Smith"),
    because the network naming guys will point out another another
    Robert.Smith.com already owns that name and you can't have it.
    
    2) The URL is really an arbitrary length base-40ish numeral.  The
    infrastructure is already in place and well tested with HTTP URLs.
    
    3) Once the URL is decoded ("Robert Smith, in Seattle, on Broad St, next to
    the donut store") it becomes a numeric address---probably a 32-bit integer
    on most systems.  Then you can address the TCP reference number plus the
    LUN to access your device.
    
    4) In your favor (although it isn't so noticable for men) names can change
    when circumstances change.  However, if you don't receive notification of
    the change, you probably shouldn't have been trying to resolve the name
    anyway.  B-)  (On a side note, the prospect of "redirectors" during login
    has been brought up, and seems very sensible, if not indispensible.)
    
    > Fibre Channel and Ethernet have taken the same approach at the
    > Port/Node WWN and MAC address level.
    
    I'm betting that the committee/person that chose 32 bits for IP wishes
    they'd/he'd chosen more bits.  Thankfully, the naming system allows almost
    unlimited expansion of the IP-bitspace behind firewalls.  People in the Bay
    Area are used to the threat of their phone numbers being changed en-masse. 
    Luckily the DNS is flexible enough to cope with (pretty much) any amount of
    oversubscription to limited numbers.
    
    > Use of a URL is a virtualization of the underlying name structure
    > and should not be implemented except as a temporary convenience for the 
    > highest level client who is constrained to a very small search
    > subset by his login authorizations.  The virtualization is always
    > at risk of security breaches.
    
    True.  But probably not significantly more so than normal.  If your secure
    e-commerce site is name-hijacked, that's a major problem too.
    
    > In the storage environment, the underlying structure is of vital
    > interest.  Any given data is present on specified physical and
    > logical units and consistency of the data is maintained on the
    > basis of an intermediate view of the lower level addressing structure
    > which must also have the proper uniqueness guarantees.
    
    This is true for the SCSI environment.  The underlying structure is
    irrelevant in the networking world.  All that matters is
    time-to-the-correct-data (or cost-to...).  In fact, it may be important to
    completely /remove/ any "intermediate views of the lower level addressing"
    because that view was, say, knocked out by a giant lizard and your current
    path to the data is being routed _around_ Tokyo now.
    
    For example, take two bridges into a legacy SAN.  There are multiple paths
    into the SAN, at different IP addresses.  One of the bridge connections is
    chewed up by a mutant space goat, so now all traffic entering the SAN has to
    pass though a bridge with a different IP address.  With a named scheme, this
    is a trivial (and automatic) repair (in fact, the DNS would mention that the
    name is associated with both IP addresses, so if one's down, use the other). 
    With an IP address, all initiators accessing that SAN would need to update
    their storage pointers.  (It's not difficult to solve, but I'm too lazy to
    want to program thousands of lists of iffy failover names.)
    
    > Address it with a URL on the browser if you wish, but understand that
    > you are talking to a data block on a single storage unit attached to
    > a network server having a specified IP address mapped to a 
    > registered MAC address.  
    
    Exactly.  Once the URL is resolved and opened, the only reference you needs
    is a TCP stream and a LUN.  (Or an SCTP connection and a LUN.)
    
    If you/we feel really strongly about this, then we could specify a maximum
    length of URL at 255 bytes (plus NUL terminator).  That would give you a
    numeric value.  (65280 bits, with admittedly lots of redundancy.)
    
    Daniel Smith.
    
    P.S., who do I call to remove mutant space goats from our data closet?
    
    P.P.S., I haven't made any arguments against using a pure numerical system. 
    But any such system will need to be of similar complexity as a URL scheme. 
    When including methods for redirection around firewalls, NATs, redundant IP
    addresses, dynamic load balancing and management side-nodes (similar to MX
    records in DNS), it gets horrendous.  My brain hurts just thinking about it. 
    B-P
    -- 
    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 01:07:00 2001
6315 messages in chronological order