Oceano

One-line summary: Server farms shared across many apps; monitoring & LB of servers, IP sprayers, and backends can be used to dynamically reallocate "fungible" servers ("dolphins") to different apps, including loading an OS+app image on the fly when a server is reallocd; fixed servers ("whales") handle heavy-duty back-end stuff (DB servers, etc) for each app.  Rule-based system is fed observed conditions and decides what to do; hysteresis and heuristics prevent extreme oscillation (eg when a server is reallocd it cannot be re-reallocd for some threshold amt of time).

Overview/Main Points

• Motivation: expanding capacity (buying/allocing/installing new HW, etc) takes O(weeks).  Bill T: IBM Olympics web farm was expanded in about 2 days since no internal bureaucracy obstacles; this may be a lower bound.
• Peaks in different services sharing a farm can be accommodate as long as they don't all peak at once.
• Key point is separation of "fixed" servers (whales) from "redeployable" ones (dolphins).

Relevance

Flaws

• How do they avoid pathological hysteresis?  If wrong assumption is made (eg if decrease in query rate is due to a failure in back end, adding more FE's may make things worse)
• It's a rule based system, but a traditional criticism of these is they become unmanageable/unpredictable when they get too big.  (Bill Tetzlaff says there's a bunch of work from the 80's where people unsuccessfully tried to apply RBS's to performance fault diagnosis)
•