Primary Storage

 View Only
  • 1.  IP replication

    Posted Wed June 24, 2020 11:36 AM
    Hi All

    When Storwize IP replication became available, IBM had a description / graphic showing the different options available that one could use regarding how many ethernet  links are supported. To be honest I could never understand that information and I currently have customers running SVC clusters with IP replication that doesn't appear to match the rules that IBM orginially specified.

    Where can I find an updated explanation of the rules around limitation on node to node IP replication?

    So as an example of I have two IO groups in production and one IO group in DR how many active links can I have?

    Alternatively  if I have a single IO group replicating to a single IO group how many active links and how many redundant links are supported?

    Any help would be much appreciated.

    ------------------------------
    Dave Buck
    ------------------------------


  • 2.  RE: IP replication

    Posted Thu July 02, 2020 12:18 PM
    Edited by Peter Basmajian Thu July 02, 2020 12:17 PM

    Hi Dave,

    There is only ever one active IP replication link per "Port Group".  So if you have two 8 node clusters, and each node has a "remote copy" IP address - the code will be picking one of the 64 possible connections and using it.

    I think we did increase the number of ports you could configure - but the number of active ports is still one per port group.

    You are only supposed to the second port group if there are two independent ethernet networks connecting the two clusters.  In that case there will be two active connections.

    So I think we allowed a lot more IP addresses to be configured - but we didn't really make any changes to how it worked.



    ------------------------------
    Andrew Martin
    ------------------------------



  • 3.  RE: IP replication

    IBM Champion
    Posted Thu September 17, 2020 02:59 PM
    One thing to also be aware of is that the latency requirements for 10 GbE ports are substantially lower than the 1 GbE.  I asked around about this some months back and it sounds like it has to do with the amount of additional information/handshakes that 10 GbE uses.  So, just keep that in mind when you're planning things out.
    • The maximum supported round-trip time between systems is 80 milliseconds (ms) for a 1 Gbps link.
    • The maximum supported round-trip time between systems is 10 milliseconds (ms) for a 10 Gbps link.


    ------------------------------
    Ian Wright
    ------------------------------



  • 4.  RE: IP replication

    Posted Fri November 13, 2020 10:21 AM
    Ian,
    The maximum supported RTT on 10GbE links is lower than on 1GbE links because of the impact of packet loss with TCP protocol. You will start decreasing your achievable throughput on a 10GbE connection at a lower RTT (basically, a lower distance), compared to a 1GbE one.

    The Mathis theory illustrates that quite clearly: 

    The Mathis Equation states that the maximum throughput achieved by a TCP connection can be calculated by dividing MSS by RTT and multiplying the result by 1 over the square root of p, where p represents the packet loss, and MSS is the maximum segment size.

    If you apply that theory to compare a 10Gb link and a 1Gb link, with the same amount of packet loss and the same segment size, you will see that the achievable throughput of a 10Gb link drops dramatically at a lower RTT than for a 1Gb connection (for which the drop of throughput appears at a higher RTT). 

    I'd say that is why we support a higher RTT on a 1Gb link (80ms vs. 10ms).

    Note: with FC protocol, we could consider there is no packet loss. Therefore, the achievable throughput is constant even over a long distance (or high RTT).

    Erwan Auffret



    ------------------------------
    Erwann Auffret
    ------------------------------