SORT BY:

LIST ORDER
THREAD
AUTHOR
SUBJECT


SEARCH

IPS HOME


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

    RE: iSCSI: Framing Steps



    
    
    > Excerpt of message (sent 29 January 2002) by Somesh Gupta:
    > > While I support a generic direct data placement model,
    > > the following are additional points for consideration for
    > > analysis of memory bandwidths and sizes.
    > > 
    > > 1. A 10G link dropping packets will not do TCP at 10Gbps.
    > > The rate drops as the packet loss increases. I don't recall
    > > but Franco or Victor from Nortel had posted an equation once.
    > 
    > A very old (40 years?) rule of thumb is that 1% loss costs you 50% in
    > throughput.  I expect that it gets a lot worse as links get faster.
    
    I would expect is to stay the same, as in your example stay 50% regardless
    of link speed. However 50% of a 1Gbps connection is ten times smaller then
    50% of 10Gbps a connection. In other words 50% of 10Gbps hurts a lot more
    then 50% 1 Gbps if you focus on the bit rate.
    
    > > ...
    > > 3. The analysis of 10Gbps, half way round the world
    > > was initially used in this debate. ALthough interesting, the person
    > > with the scenario above is not going to blink at the cost of
    > > 256MBytes of fastest memory considering what they are paying
    > > for the link.
    > 
    > Exactly.
    
    Yup. I believe the concern about high-speed connections and longer round
    trip times can be solved by upping the TCP window and correspondingly upping
    the resource requirements (mostly additional RAM) to deal with it. This
    could be avoided by not using TCP as it exists today but the issues with
    that are many (already heavily discussed) and can IMHO outstrip the costs of
    additional RAM. In general I think the costs of RAM are dropping far faster
    then the costs of upgrading the network.
    
    > One reason ATM failed as a LAN is that its design was burdened with
    > complexity based on that sort of scenario.  (In other words: "it has
    > to work at umpteen Gig, across the globe, and only use a tiny little
    > bit of memory because implementations can't handle a megabyte of
    > buffers".) 
    > 
    > A design optimized for the combination of very high bit rate and very
    > long latency will inevitably be way overpriced for the predominant
    > case, which is the LAN case.
    
    I would tend to agree.
    
    -Shawn
    
    p.s. I am not speaking for Hewlett-Packard this is just my opinion.
    


Home

Last updated: Tue Jan 29 17:17:59 2002
8551 messages in chronological order