Scalable concurrency control in a dynamic membership
Download as ODP, PDF0 likes298 views
The document summarizes research on scalable concurrency control in dynamic distributed systems using a multi-token approach. The approach proposes using a mesh overlay topology and random routing of tokens to control access to a shared resource among a large number of dynamic nodes. Experimental results showed the process converges quickly but with more tokens and worse performance than expected, requiring further tuning of the control loop dynamics.
Scalable concurrency control in a dynamic membership
- 1. Scalable concurrency control in a dynamic membership A. Ciuffoletti * – A. Asim ** * INFN-CNAF, Bologna, Italy ** University of Paris Sud-XI, Paris, France 3 rd CoreGRID Workshop on Middleware – June 2008
- 2. Problem statement A centralized resource A large, dynamic number of users Each user must be periodically granted access to the resource A use case A centralized registry that must be periodically checked by each user to synchronize a local cache.
- 3. A multi-token approach Given The service time (here token latency ) t and The number of nodes in the system N We can figure out the number of tokens needed to sustain a certain period p : But we need: An overlay ring (difficult) The size of the network (difficult)
- 4. Our solution Overlay topology: mesh with degree near to N Token routing: random neighbor Token number: closed loop over p Problems: Security issues (tokens are received from “anybody”) Maintain a cache with most of the neighbors Closed loop control
- 5. Closed loop control Measure the time from the last token receive event Early tokens are delayed in a short queue (discard on overflow) Generate a new token upon timeout Appear to be quite simplistic: loop control is still to refine Problem (minor): sensitivity to network performance
- 6. Membership maintenance Recent updates to the local membership directory are piggybacked to the token Limited capacity of the token At each step the list of updates is recomputed Join event Create a new token and send it to a known peer Leave event File a leave event in the first availabletoken
- 7. Experimental results Decrease exponentially (good) Maximum at the minimum (bad) Average at 1.7 instead of 2.6 (bad) Waving? Due to buffering.
- 8. Experimental results The process converges (good!) Its dynamics are extremely fast (unexpectedly good) Number of token exceeds expectation (55 instead of 2!)
- 9. Experimental results All join events eventually received (good) Takes a long time (expected)
- 10. Conclusions One of the first experiments of this kind in the Internet scale (instead of theoretical investigation and simulations) Mixed results: The process stabilizes, local directories eventually converge Dynamics are not those expected: too many tokens, short lived, too frequent Overall: we are on the right way, tuning needed