Preview of Poster and Demo session

1
Preview of Poster and Demo session

• New Brighten Room
• 730 PM to 830 PM Tonight

2
Scaling Ruby on Rails

• Understand performance of Ruby on Rails
  applications on modern manycore processors (e.g.
  Sun Niagara 2) and cloud computing environments
• Identify scaling bottlenecks in apps and other
  runtime elements (middleware, web server, etc.)
• Develop best practices for deployment,
  development, monitoring, etc. for scalability
• Approach
• Port existing Sun Web 2.0 example application
  (social networking) to Ruby on Rails
• Use Faban to exercise application with varying
  load
• Collect data and analyze data from runs

Will Sobel, Arthur Klepchukov, Hubert Wong
3
Evaluating Amazons EC2 As a Research Platform

• EC2 is appealing
• 1000 machines for only 100
• Concerns
• Unknown hardware under virtualization
• This project
• Characterize the performance and variance of
  different aspects of EC2
• Provide recommendations

Michael Armbrust Gunho Lee
4
Ruckus

• Insatiable need for data
• Needs systematic solution
• Get the data we need to where we need it.
• Do it at scale
• Do it cheaply
• Ruckus aims to achieve these goals via
  declarative techniques.

Ari Rabkin
4
5
Mining Text Logs to Detect Server Problems

• Automatically analyze text logs without
  specifying query
• Three Case Studies
• Suns Project Darkstar
• Hadoop
• A production distributed storage system
• Detect hard-to-notice problems
• One extra problem case in poster session only!

Wei Xu
6
SCADS

• Motivation
• Web frameworks make it very easy for programmers
  to design compelling applications that serve
  millions of people
• Successful applications quickly discover
  scalability limitations of traditional RDBMS
• The result is complicated ad-hoc infrastructures
  on top
• Goal
• Create a structured data storage system for
  interactive web applications, designed from
  beginning to scale.
• Allow developers to easily reason about
  consistency/performance tradeoffs

Michael Armbrust Beth Trushkowsky
7
Improving MapReduce Performance in Large
Virtualized Environments

• MapReduce is becoming a popular model for
  large-scale computation
• EC2 provides cheap CPU power, at the cost of
  virtualized environment
• Evaluated Hadoop MapReduce on EC2, collecting
  data with X-Trace
• Identified heterogeneity as key problem in
  virtualized environment
• Designed heterogeneity-aware scheduler that
  improves performance up to 2x

Matei Zaharia Andy Konwinski
8
Understanding Performance Variability in MSN
Messenger

• Whats causing high variance of latency of MSN
  Messenger?
• under steady, test workload
• Approach
• which section of execution has most variance?
• using request paths
• which metrics correlate with high variance?

Peter Bodik
9
Characterizing Workload and Provisioning for
Scale Up

• Motivation
• EC2 Workload Scheduling
• Predicting scale up performance
• Methodology
• Linear regression to understand job-specific
  scaling characteristics
• Predict performance using regression curves from
  micro-benchmarks

Interesting Image Or Graph from Poster
Archana Ganapathi
Photo of you
10
AWE-Sim Towards a Realistic and Privacy
Preserving Simulator for Proprietary Systems

• Realistic workload generation needed for
  research.
• Project explains how model that abstracts from
  raw data may be used to simulate system behavior.

Kristal Sauer
Archana Ganapathi
11
Diagnosing Faults Using Queueing Networks
Graphical Models

• Broken aspects of queueing networks
• Strong distributional assumptions
• Time-dependent statistics difficult

FIX ML algorithms for inference in probability
distributions
Gibbs sampling EM
Charles Sutton, George Porter, Randy Katz, and
Michael I. Jordan
12
Dynamic Middleboxes -What, Why, When, Where and
hoW

• Middleboxes like firewalls and load balancers are
  becoming virtualized and dynamic.
• Challenges
• How to get traffic to dynamic middleboxes?
• Policy-aware Switching Layer
• When to turn on/off?
• Where to place?
• Will they scale?
• Middlebox virtualization platform?

Load balancer
firewall
Servers
Dilip Joseph
13
Emulating 10,000 Servers with 20 FPGAs

• SPARC v8 32-bit target nodes, running unmodified
  binaries
• Emulation scale
• 512 nodes/FPGA, 2048 nodes/BEE3 board, 10,240
  nodes with 20 FPGA
• DRAM capacity
• 16 GB/FPGA, 32 MB/node (64 contexts/CPU)
• Emulation Performance
• Will run at 150 MHz on Xilinx Virtex 5 LX110
• gt 1 GIPS/FPGA

Zhangxi Tan
14
Evaluating Reliability, Availability and
Serviceability (RAS) Capabilities

• Model and measurement driven evaluation approach
• Use runtime adaptation to inject faults or induce
  failures in live systems
• Use analytical model templates to identify
  weak-points, construct failure scenarios and
  score responses

Rean Griffith
15
Deterministic Replay of Multicore Applications
Logging Overhead

• Goal reproduce bugs in datacenter by replaying
  apps
• Key obstacle logging overhead on multi-cores is
  impractically high
• Best known method log all reads from memory
• Our method infer the values of reads from memory

Gautam Altekar
16
Virtics Isolating Malware on the Desktop
Can we rid the desktop of malware if we are
smarter about partitioning our systems?
Virtics sets out to do exactly that by running
every program and opening every document in its
own virtual machine.
Matt Piotrowski
17
CT-NOR Modeling Network Dependencies

• In a network, dependencies are hard to understand
• A services failure leads to more failures
  downstream
• Need automated tools
• CT-NOR - statistical model that finds
  dependencies
• Based on packet timing only
• Protocol-agnostic
• Models the distribution of output packets using
  Poisson Processes
• Used for Constellation project at MSR

Alex Simma
Photo of you