1. Improving Robustness in Distributed Systems Per Bergqvist per@synapse.se per@synapse.se Erlang User Conference 2001 (courtesy CellPoint Systems AB)

2. Design base Cluster of cooperating hosts Erlang and C COTS hardware based Unix based (i.e. Solaris or Linux) 10/100/1000 base-T back plane (”system area network”)

3. Cluster Shared, distributed, system configuration Each host have ONE cluster controller Dispatch and supervise worker tasks Master cluster controller: holds configuration database (persistent replica) Slave cluster controller: gets configuration from master cluster controllers Cluster is DOWN when all master cluster controllers are inaccessible

4. Typical system Firewall Switch Traffic Control

5. Cluster Key Benefits Single system view Enforces decoupling of parts of O&M from actual traffic processing

6. Implementing a cluster Cluster->Host->Node->NodeData Cluster global parameters Subscription mechanisms for conf. changes Mnesia as configuration database on master cluster controllers Homebrewn configuration distribution to slave controllers (NOT using mnesia) (Worker) node supervision

7. Mnesia gotchas First distributed node startup Disallow writes when all replicas not accessible Use timeout on table load and force load

8. ... BUT... TCP based distribution Network partitioning

9. Network parameters Align TCP retransmission intervals w/ Erlang heartbeats Align TCP and IP rerouting parameters

10. Typical system II: Dual back plane Firewall Switch Traffic Control

11. Erlang multi-homing problem Host A Host B Host C

12. Multi-home Erlang w/ TCP Add an alias interface to loop back i/f Patch tcp distribution to bind to alias Publish alias interface on (all wanted) via real hw i/f’s Method 1: Static routes and gratuitous/proxy arp Method 2: Use new (routing) protocol

13. ARP method Implement a utility to: - broadcast unsolicited ARP responses - respond to ARP requests for the alias i/f address Add static routes on all far end systems NOTE: all real i/f needs to be on same IP subnet

14. New routing protocol Broadcast (Ethernet frames) what you have, including interface priority Let the far end select path based on what/when they receive Far end dynamically sets up host routes Use short retransmission intervals

15. Erlang multi-homing resolved ? Host A Host B Host C

16. Summing up Erlang can support multihoming with some additional work By using loop back alias i/f, link failure becomes a routing problem (peer-peer association is kept intact) Solaris TCP/IP stack parameters are: - hard to find (only in out-of-date app. notes) - hard to set ”right” - host global A distribution mechanism with built-in support for multi-homing preferred

17. Erlang Distribution over SCTP Per Bergqvist et al per@synapse.se per@synapse.se Erlang User Conference 2002