Transcript Towards an Agile, Predictive Infrastructure

Griffin Update: Toward an Agile,
Predictive Infrastructure
Anthony D. Joseph
UC Berkeley
http://www.cs.berkeley.edu/~adj/
Sahara Retreat, June 2003
Outline

Griffin
–
–
–



2
Motivation
Goals
Components
Tapas Update
Tapestry/Brocade Update
REAP/MINO Update
Near-Continuous, Highly-Variable
Internet Connectivity

Connectivity everywhere: campus, in-building, satellite…
–

Most applications support limited variability (1% to 2x)
–
–

–
–
–
–

Design environment for legacy apps is static desktop LAN
Strong abstraction boundaries (APIs) hide the # of RPCs
But, today’s apps see a wider range of variability
–
3
Projects: Sahara (01-), Iceberg (98-01), Rover (95-97)
35 orders of magnitude of bandwidth from 10's Kb/s 1 Gb/s
46 orders of magnitude of latency from 1 sec 1,000's ms
59 orders of magnitude of loss rates from 10-3  10-12 BER
Neither best-effort or unbounded retransmission may be ideal
Also, overloaded servers / limited resources on mobile devices
Result: Poor/variable performance from legacy apps
Griffin Goals

Users always see excellent ( local, lightly loaded)
application behavior and performance
–
–
–

Help legacy applications handle changing conditions
–
–

Analyze, classify, and predict behavior
Pre-stage dynamic/static code/data (activate on demand)
Architecture for developing new applications
–
–
4
Independent of the current infrastructure conditions
Move away from “reactive to change” model
Agility: key metric is time to react and adapt
Input/control mechanisms for new applications
Application developer tools
Griffin: An Adaptive, Predictive
Approach

Continuous, cross-layer, multi-timescale introspection
–
–
–

Convey app reqs/network info to/from lower-levels
–
–

Break abstraction boundaries in a controlled way
Challenge: Extensible interfaces to avoid existing least
common denominator problems
Overlay more powerful network model on top of IP
–
–
5
Collect & cluster link, network, and application protocol events
Broader-scale: Correlate AND communicate short-/long-term
events and effects at multiple levels (breaks abstractions)
Challenge: Building accurate models of correlated events
–
Avoid standardization delays/inertia
Enables dynamic service placement
Challenge: Efficient interoperation with IP routing policies
Some Enabling Infrastructure
Components

Tapas network characteristics toolkit
–
–
–

REAP protocol modifying / application building toolkit
–
–
–

Introspective mobile code/data support for legacy / new apps
Provides dynamic placement of data and service components
MINO E-mail application on OceanStore / Planet Lab
Tapestry, Brocade, and Mobile Tapestry
–
6
Measuring/modeling/emulating/predicting delay, loss, …
Provides micro-scale network weather information
Mechanism for monitoring/predicting available QoS
–
Overlay routing layer providing efficient application-level object
location and routing
Mobility support, fault-tolerance, varying delivery semantics
Outline

Griffin
–
–
–



7
Motivation
Goals
Components
Tapas Update
Tapestry/Brocade Update
REAP/MINO Update
Tapas

Accurate modeling and emulation for protocol design
–
–
–

Goal: Create models/artificial traces that are
statistically indistinguishable from real network traces
–
–
–

Such models have both predictive and descriptive power
Better understanding of network characteristics
Can be used to optimize new and existing protocols
Tapas: Novel data preconditioning-based analysis
–
8
Very difficult to gain access to new or experimental networks
Delay, error, congestion in IP, GSM, GPRS, 1xRTT, 802.11a/b
Study interactions between protocols at different levels
More accurately models/emulates long-/short-term
dependence effects than classic approaches (Gilbert, Markov,
HMM, Bernoulli)
Tapas Update

Domain analysis tool
–
Chooses most accurate model for a metric


New Tapas-based link simulator
–
–

9
Markov-based Trace Analysis, Modified hidden Markov Model
Complete reimplementation of Wsim
Enables quick and repeatable testing of new apps
Tapas talk this afternoon
Domain Analysis:
Choosing the Right Network Model

Collect empirical packet trace: T = {1,0}*
–

1: corrupted/delayed packet, 0: correct/non-delayed packet
Create mathematical models based on T
real network
metric trace

–

artificial network
metric trace
Gilbert, HMM, MTA, M3 have different properties
Algorithm (applied to Gilbert, HMM, MTA, M3):
–
10
network
model
Challenge: domain analysis – which model to use?
–

trace analysis
algorithm
Collect traces, compute exponential functions for lengths of good
and bad state and compute 1’s density of bad state
For a given density, determine model parameters and optimal
model (best Correlation Coefficient)
Sigmetrics 2003 paper
Domain Analysis Experiment


Create artificial network environment with varying bad
state densities (generate synthetic reference traces)
For each trace:
–
–
–
–

Optimal model for a given set of properties is the one with
the highest CC value
–
11
Create classical and data preconditioning models
Generate artificial traces from models
Plot error and error free distribution
Calculate Correlation Coefficient (CC) between distributions of
reference and artificial traces
Plot optimal model as a function of the good and bad state
exponential values: Domain of Applicability Plot
Domain of Applicability Plot, Lden= 0.2
12
Domain of Applicability Plot , Lden= 0.7
13
Tapestry/Brocade

Starting point is Tapestry
–

Extend Tapestry with unique, powerful routing functions
–
–
–
–
14
Distributed Object Location and Routing (DOLR) overlay network
SLA-compatible efficient wide-area routing
Rapid, scalable mobility support
Rapid fault route-around using pre-computed backup routes
Monitoring, measurement, and analysis entry point
Tapestry/Brocade Update

Major push to improve Tapestry reliability
–
–
Pre-computed backup paths enable nearinstantaneous fail-over (3 paths/router entry)
Improved Patchwork network link monitoring

–

15
Improved repair algorithms to handle long-term faults
Building new applications
–

Now ready for integration with link prediction support
SpamWatch (Middleware 2003 paper)
Summer focus on inter-domain Brocade routing
Improved Tapestry Fault-Tolerance
16
REAP/MINO

Introspective code / data migration in 3-tier hierarchies
–
–

Combines static trace analysis w/ dynamic monitoring
of clients to predict appl’n / communication behavior
–
–
–
–

Identify and optimize code/data placement
Pre-stage statically/dynamically generated components
Explore various granularities of code & data migration
Predict costs using multiple criteria
MINO E-mail OceanStore application
–
17
Distributes server load, empowers limited devices
Provides illusion of high connectivity
Basis for exploring code/data migration choices
REAP/MINO Update

Code migration work is mostly complete, now focused
on data migration
–

Collecting user mobility / data access traces
–
–
–

Web proxy traces from NLANR (access clustering)
EECS IMAP server traces (user location, access info)
College campus NFS-level traces from a login server (used for
e-mail reading)
Built cooperative caching simulator
–
18
Understanding how users access data and how they move, so
that we can better place/cache data
–
Exploring multi-criteria optimization: frequency of access,
number of users accessing, $ paid/user, etc..
Cache refill can leverage link predictor information
Recent Griffin Progress Summary

Tapas: Network modeling and analysis
–
–
–
Thesis: Almudena Konrad, “TAPAS: A Research Paradigm for the
Modeling, Prediction and Analysis of Non-stationary Network
Behavior,” (PhD, expected August 2003)
Simulator almost ready for release
Publications


19
Konrad, A.; Joseph, A. D., Choosing an Accurate Network Path
Model, In Proc. Of SIGMETRICS 2003, June, 2003.
Konrad, A.; Zhao, B. Y.; Joseph, A. D.; Ludwig, R., A Markov-Based
Channel Model Algorithm for Wireless Networks, ACM Wireless
Networks, vol. 9, num. 3, May, 2003.
Recent Griffin Progress Summary

Tapastry / Brocade:
–
–
Robustness fixes to Tapestry, lots of measurements
Publications



REAP/MINO
–
–
20
Zhao, B. Y.; Huang, L.; Stribling, J.; Rhea, S. C.; Joseph, A. D.;
Kubiatowicz J. D., Tapestry: A Resilient Global-scale Overlay for
Service Deployment. To appear in IEEE JSAC, Fall 2003.
Zhou, F.; Zhuang, L.; Zhao, B.; Huang, L.; Joseph, A. D.;
Kubiatowicz, J., Approximate Object Location and Spam Filtering
on Peer-to-Peer Systems, In Proc.of ACM Middleware 2003,
June, 2003.
Simulator developed, lots of traces collected
Beginning analysis phase
Griffin Update: Toward an Agile,
Predictive Infrastructure
Anthony D. Joseph
UC Berkeley
http://www.cs.berkeley.edu/~adj/
Sahara Retreat, June 2003