[email protected] (John McKown) writes: > Back in the z890 days, we had a CPU fail. Of course, the hardware > automatically recovered and we only knew about it due to a logrec record > being written and a message on the HMC. We also had one of our OSAs fail. > The second OSA did an ARP takeover (proper term?) and we suffered _no_ user > interruption. The LAN people _refused_ to believe that the OSA could fail > that way without disrupting all the IP sessions of the users on that OSA. > Apparently when a multi-NIC PC has a NIC fail, all the IP sessions on that > NIC die immediately (well, they time out).
re: http://www.garlic.com/~lynn/2014.html#23 Scary Sysprogs and educating those 'kids' http://www.garlic.com/~lynn/2014.html#24 Scary Sysprogs and educating those 'kids' http://www.garlic.com/~lynn/2014.html#27 Hardware failures (was Re: Scary Sysprogs ...) we did IP-address take-over (ARP cache times out and remaps ip-address to a different MAC address) in HA/CMP http://www.garlic.com/~lynn/subtopic.html#hacmp however at the time, most vendors used bsd reno/tahoe 4.3 software for their tcp/ip stack ... and there was a "bug" in the 4.3 code (and therefor in nearly every machine out there). the bug was in the ip layer ... it saved the previous response from call to ARP cache ... and if the next IP operation was for the same ip-address, it used the saved value (and bypassed calling arp cache handler). ARP cache protocol requires that the saved ip-address/mac-address mapping in the ARP cache times-out and a new ARP operation has to be done to discover the corresponding MAC address (for that ip-address). However, the "saved" mac address had no such time-out. In a strongly oriented client/server environment when the client primarily does majority of its tcp/ip to the same server (ip-address) ... it could go for long periods of time w/o changing ip-addresses. As a result a server doing ip-address takeover to a different LAN/MAC address wouldn't be noticed by such clients. We had to come up with all sorts of hacks to smear ip-address traffic across the environment ... trying to force clients to reset their ip-address to mac-address mapping. There is separate gimmick which involves MAC-address spoofing ... i.e. in theory every MAC-addresses are unique created at manufacturing time ... however some number of adapters have been given the ability to soft reset their MAC-address (so if one adapter fails ... another adapter can spoof the failed adapter). -- virtualization experience starting Jan1968, online at home since Mar1970 ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [email protected] with the message: INFO IBM-MAIN
