1. Computational Risk Management for Building Highly Reliable Network Services Chaki Ng Brent N. Chun Philip Buonadonna HotDep’05

2. Chaki Ng || Computational Risk Management 2 Network Service Performance Desire for Hard Performance Guarantees  “99.999% availability,” “all trades < 30 seconds” Difficult to Achieve Consistently  Demand: workload varies and can be bursty  Supply: resource needs vary and hard to plan for Dedicated and Over-Provisioning  $$$, low utilization Shared Infrastructure  Resource supply varies – competition, failures Tradeoff supply and performance guarantees

3. Chaki Ng || Computational Risk Management 3 Computational Service Provider (CSP) Goal: mechanism to manage supply Resources (e.g. server nodes)  Accommodate peak demand of most services Markets of nodes  Each node sells resource contracts Spot, futures, options  Contracts priced based on supply and demand

4. Chaki Ng || Computational Risk Management 4 Measure Risk How to quantify performance guarantees  Risk metrics: simple statistical summaries of undesirable outcomes Example: Value-at-Risk (VaR)  Finance: “The Fidelity mutual fund will lose no more than $25MM monthly, with 95% probability”  Computation: “Amazon.com will process orders in less than 30 seconds daily for 95% of all orders” Two challenges: calculate VaR and sensitivity analysis of VaR

5. Chaki Ng || Computational Risk Management 5 Calculate VaR Calc expected performance distribution Example method: historical Methods: Variance, Monte Carlo, Stress Testing Probability Fidelity Fund Profit/Loss 95% Var: -$27MM Probability Amazon.com Order Time 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 95% Var: 33 seconds

6. Chaki Ng || Computational Risk Management 6 Compute VaR: Model Supply and Demand Own Service Workload Forecast Node Performance and Trade Forecast Aggregate Workload Forecast VaR Set of Accessible Node Resources Supply

7. Chaki Ng || Computational Risk Management 7 Sensitivity Analysis of VaR Goal: model how VaR varies as the set of resource contracts changes  VaR = F(set of resource contracts) Forecast demand and supply  Nodes and aggregate workload forecast  Own client workload forecast Model portfolio VaR  Swap set of resource contracts  Calculate VaR improvements

8. Chaki Ng || Computational Risk Management 8 Portfolio Management Goal: meet target VaR within budget and minimal cost Continuous portfolio optimization  Find available set of resources  Find sets that achieve best VaR Trade resource contracts  Buy best set within budget

9. Chaki Ng || Computational Risk Management 9 Finance: Manage Portfolio VaR Portfolio VaR Target VaR: “The Fidelity mutual fund will lose no more than $25MM monthly with 95% probability.” IBM MSFTORCL Probability Fidelity Profit/Loss EBAY Financial Markets Sell IBM @ $75 Buy EBay @ $37 95% Var: -$27MM

10. Chaki Ng || Computational Risk Management 10 Probability Amazon.com Order Time 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 95% Var: 33 seconds Node4 Computation: Manage Portfolio VaR Portfolio VaR Target VaR: “Amazon.com will process orders in less than 30 seconds for 95% of all orders.” Node1 Node2Node3 CSP Sell Node1 @ $50 Buy Node4 @ $30

11. Chaki Ng || Computational Risk Management 11 Open Problems Resource Contracts: pricing, base units Programming: model, API Modeling Supply and Demand Portfolio Strategies: “standard portfolios” Interoperability: across different CSPes

12. Chaki Ng || Computational Risk Management 12 Conclusion Dedicated vs. shared CSP: share resources via markets Achieve performance goals in the context of shared CSP  Quantify performance goal via risk metrics like VaR  Calculation and sensitivity analysis  Portfolio optimization

13. Chaki Ng || Computational Risk Management 13 Backup Slides

14. Chaki Ng || Computational Risk Management 14 Simple Experiment Service Workload Node Failures Each request tries N nodes randomly If both nodes down  failed request Daily Service Availability = Failover  Successful Requests  All Requests

15. Chaki Ng || Computational Risk Management 15 Results Each point: 100 daily runs, 100 requests/hr