Selected Papers in CCGrid 2004 Presented by Chan

Selected Papers in CCGrid 2004 Presented by Chan Chi Yuk

1. Data Durability in Peer to Peer Storage System § Authors: § Gil Utard and Antoine Vernois § In the paper: § Replication vs. Erasure code § (PAST, Freenet) vs. (Ocean. Store) § Inter. Memory § Volatility factor and Availability factor § Markov Chain model and system MTTF

Node Behaviour § § § § State u: temporarily unavailable or disconnected State c: connected State d: permanently unavailable or dead Volatility factor: : average lifetime of state c, availability factor : average lifetime of u + =1

State transition for the system § § Red: failure Green: connection Brown: disconnection Blue: repair § (time for repair = 1) § Non recurrent Markov Chain with one absorbing state

MTTF from calculation! =0. 1 =0. 3

Why have such MTTF values?

What the paper concludes § Erasure codes require high availability of peers to be efficient § They have also shown data recover cost is very high for low availability (not covered) § Future work: § Mixes replication and erasure code

2. Analysis of Large-Scale Topological properties for Peer-to-Peer Networks § Authors: § Hao Chen, Hai Jin, Jianhua Sun, Dafu Deng, and Xiaofei Liao § The topology of Gnutella § § § By crawler: crawler ping By web spider: get information page Discover 50, 000 peers within 0. 5 hr (~50%) § Their Study: § § Small world Power law Centrality Modeling

Cluster coefficient decays with higher node degree According to the authors, no obvious evidence for power law

Degree distribution, power law?

What you may also be interested: Centrality Degree Centrality Closeness Centrality Betweeness Centrality Network Centrality

3. Adaptive Multi-Resource Prediction in Distributed Resource Sharing Environment § Authors: § Jin Liang, Klara Nahrstedt, and Yuanyuan Zhou § Related works: § LAST § EWMA (exponential weighted moving average) § AR (autoregressive) § MModel(p, q) (Multi-Resource Prediction Model) ACF: XCF:

Adaptation techniques § MModel-a § Mean adaptation § =0. 99 § XCF adaptation § q=1: limit the potential benefit of cross correlation, but simplify the adaptation § 1=0. 9: respond more quickly to cross correlation

Two sets of Traces § 1 min CPU load average and free memory on 5 machines in the CS Instructional Lab at UIUC § 12 machines of the NCSA Origin 2000 supercomputer arrays

Results: MModel-a is good

Results: Not good for all cases

Other Papers § Sizing the Streaming Media Cluster Solution for a Given Load § Remote Partial File Access Using Compact Pattern Description § An Efficient Replicated Data Access Approach for Large-Scale Distributed Systems