ericvaandering
commented
10 years ago Comment by egeland on Fri Jan 23 08:53:49 2009
No longer part of our plans.
Closed ericvaandering closed 10 years ago
ericvaandering
commented
10 years ago Comment by egeland on Fri Jan 23 08:53:49 2009
No longer part of our plans.
ericvaandering
commented
10 years ago Closed by egeland on Fri Jan 23 08:53:49 2009
Original Savannah ticket 1260 reported by None on Mon Oct 18 15:12:13 2004.
Ideas from CHEP 2004 / Interlaken / 27.9.2004
0) Obtain topology server list (default to built-in list) 1) Get topology -> returns routing table with time stamps and TNDM (Tier-N data matrix) contact; to be called on all configured servers: well-known (central?) service and all P2P-configured services 2) Figure out nearest servers based on user configuration (e.g. belonging to Tier-N as Tier-N+1) 3) Obtain data matrix with time-stamps => find out who's got what (multi-level data -- get only details for data that was subscribed; stream/dataset vertical slices); may need to request data matrix from many servers (subscription) or from parent (sub-subscription) -> sub-subscription when you only want subset of data at parent node, subscription when you want data no matter where it is 4) Request list of replicas from TNRT (Tier-N replica table) -> insert into own database as "replica frontier" 5) Create transfer request -> run own allocator 6) Create send requests behind frontier (sending server)
Data is multi-level: you only get what you are interested in. When get interested, expand it into local database from peers.
Server may refuse to talk to you, or create transfer requests for you, e.g. T1 may choose to only serve T2 nodes.
All data is time-stamped. Need to figure out efficient methods for syncing data between e.g. replica tables. Could be enough to get own max(timestamp) and ask server to send everything newer than that.
Examine P2P systems and IP routing algorithgms to determine efficient way of "getting plugged in": discovering services -> discovering toplogy -> discovering data -> replicating it. Preferably just "get software, start it and go", will automatically discover services and topology. Probbaly need to be told about some (VO?) central services, can be configured in software distribution.