1. Thomas Dreibholz Institute for Experimental Mathematics University of Duisburg-Essen, Germany dreibh@iem.uni-due.de University of Duisburg-Essen, Institute for Experimental Mathematics A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments

2. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 2 Table of Contents What is Reliable Server Pooling?  Prototype Demonstration  Terminology and Protocols  Motivation and Application Scenarios The Challenge on Network Delay on Server Selection The Least Used with Degradation Policy Evaluation Conclusion and Outlook Thomas Dreibholz's Reliable Server Pooling Page http://tdrwww.iem.uni-due.de/dreibholz/rserpool/ Thomas Dreibholz's Reliable Server Pooling Page http://tdrwww.iem.uni-due.de/dreibholz/rserpool/

3. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 3 What is „Reliable Server Pooling“? Prototype Demonstration

4. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 4 Reliable Server Pooling (RSerPool) Terminology:  Pool Element (PE):Server  Pool:Set of PEs  PE ID:ID of a PE in a pool  Pool Handle:Unique pool ID  Handlespace:Set of pools  Pool Registrar (PR)  Pool User (PU):Client  Support for Existing Applications Proxy Pool User (PPU) Proxy Pool Element (PPE) Protocols:  ASAP (Aggregate Server Access Protocol)  ENRP (Endpoint Handlespace Redundancy Protocol)

5. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 5 What is a Pool Policy?  A rule for the selection of the PEs  Defined in our IETF Working Group draft (draft-ietf-rserpool-policies-07.txt) Application of Policies  Registrar: Creates PE list upon request by PU  Pool User: Selection of a PE from the list  Both according to the pool policies (pool-specific!) Non-Adaptive Policies  Stateless: Random (RAND)  Stateful: Round Robin (RR)(Default policy, must be supported) Adaptive Policy  Least Used (LU) Load definition is application-specific! Round robin among multiple least-loaded PEs Server Selection Rules (Pool Policies)

6. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 6 The Challenge of Network Delay on Server Selection Challenge of Least Used  Load states get out of date, due to Network latency Cache Solution: Least Used with Degradation (LUD)  Policy Information: Load = Current Load (obvious) Load Increment = How much is load increased by a new request?  Select PE, which has lowest sum of (Load + Load Increment)  Round robin among equal-valued PEs  Upon selection: Increment load by load increment Incrementation only local on selection component (i.e. registrar and PU's cache)!  Upon update: Load is reset to latest known load state

7. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 7 The Application Model Server –PE Capacity –Shared among sessions (multi-tasking principle) Client –Requests are generated Request Size (effort) Request Interval (frequency) –Waiting queue for requests –Sequential processing System Utilization –PU:PE Ratio –Provisioning for certain Target Utilization, e.g. 80%

8. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 8 Performance Metrics Provider's Perspective “Does my server capacity gain revenue?” Average Utilization of server resources [%] User's Perspective “How much time is needed to process my requests?”  Avg. Handling Speed [% of average server capacity]  Depends on: Queuing Startup Server

9. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 9 Increasing the Network Delay - A Proof of Concept Example setup as a proof of concept Network latency reduces the handling speed...... but with LUD, there is a significant speed benefit compared to LU More investigations necessary  Workload parameters  Number of registrars  Cache Handling Speed

10. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 10 Variation of Workload Parameters: PU:PE Ratio Small PU:PE ratio is critical (high per-PU workload) LUD achieves significant performance improvement over LU Handling Speed Utilization LU, Req.Int=10s (critical!)

11. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 11 Variation of Workload Parameters: Request Interval Small request interval is critical (especially for small PU:PE ratio!) For PU:PE ratio > 1, LUD again achieves a significant improvement Handling Speed Utilization

12. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 12 Increasing the Number of Registrars Handlespace synchronization  Necessary to cope with PR failures  Additional load update latency Results:  LUD again achieves a significant benefit over LU... ... for realistic number of PRs (less than 10) Handling Speed

13. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 13 Using the PU-Side Cache Cache at the PU:  Stores partial, temporary subset of the handlespace  Reduces number of PR queries  Contents get out of date Results:  Again, LUD outperforms LU Handling Speed

14. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 14 Conclusion and Outlook Conclusion  RSerPool is the IETF's upcoming standard for service availability  Network delay leads to out-of-date load states for Least Used policy  Least Used with Degradation (LUD) Local increment upon selection, until update arrives Improved system performance, especially for critical workload parameter settings Future Work  From simulation to reality: Tests with our prototype implementation in the PlanetLab First results already available [KiVS2007]  Security analysis and robustness against DoS attacks

15. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 15 Thank You for Your Attention! Any Questions? Visit Our Project Homepage: http://tdrwww.iem.uni-due.de/dreibholz/rserpool/ Thomas Dreibholz, dreibh@iem.uni-due.dedreibh@iem.uni-due.de To be continued...

16. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 16 The RSerPool Protocol Stack Aggregate Server Access Protocol (ASAP)  PR  PE: Registration, Deregistration and Monitoring by Home-PR (PR-H)  PR  PU: Server Selection, Failure Reports Endpoint Handlespace Redundancy Protocol (ENRP)  PR  PR: Handlespace Synchronisation ASAP is IETF's first Session Layer standard! ASAP is IETF's first Session Layer standard!

17. Thomas Dreibholz A New Server Selection Strategy for Reliable Server Pooling in Widely Distributed Environments P. 17 Motivation Motivation of RSerPool:  Unified, application-independent solution for service availability  Not available before => Foundation of the IETF RSerPool Working Group Application Scenarios for RSerPool:  Main motivation: Telephone Signalling (SS7) over IP  Under discussion by the IETF: Load Balancing Voice over IP (VoIP) with SIP IP Flow Information Export (IPFIX) ... and many more! Requirements for RSerPool:  “Lightweight” (low resource requirements, e.g. embedded devices!)  Real-Time (quick failover)  Scalability (e.g. to large (corporate) networks)  Extensibility (e.g. by new server selection rules)  Simple (automatic configuration: “just turn on, and it works!”)