Mapping Marin

1
Mapping Marins Failure Detectors to JuxMems
LDGs/GDGs

• Sébastien Monnet
• GDS Meeting
• May 4th, Rennes

2
LDG / GDG
3
Failure detectors hierarchy
4
Mapping

• Intra-cluster
• FD local / LAN
• JuxMem
• LDG
• Cluster group
• Inter-cluster
• FD global (leader to leader)
• JuxMem GDG (leader to leader)
• Notion of leader in both

5
Software architecture

• Does not perfectly fit to the current
  architecture
• No hierarchy knowledge at FDs level
• 2 main solutions
• Bind the architecture to fit
• Build a FD service that monitors all the JuxMem
  peers

6
Bind the architecture to fitSolution A

• Interface between FTGM and FD
• Give to FD peers IDs and group IDs separately
• It becomes hierarchy aware

7
FD serviceSolution B

• FD launch on each JuxMem peer at star time
• A FD stub register to the FD service
• Filter unmonitored peers

Failure detection service
8
Implementation issues

• In both cases
• Mapping IPport / ID
• IP multicast
• JXTA propagate pipes
• Retrieving via JXTA all needed information
  (detecteur.ini)
• Solution A
• Unify the notion of leader
• Solution B
• Interface between FTGM and the FD service
• Registration
• Notifications

9
Performances issues

• Solution A
• A single peer may belong to n LDG
• Multiple monitoring
• Solution B
• Few peer may provide data
• Useless monitoring of certain peers (e.g.
  clients)
• Launch FD only on interesting peers (on the fly) ?

10
Solution A not so bad
Solution A
Solution B
11
Ongoing work

• Done
• FD sources transfer (Marin -gt Sebastien)
• Mini tutorial (in Chicago, IL )
• To do
• Still need some explanations
• Implementation / integration
• A paper