How We Use the NSFCloud to
Save the World
Yan Luo
Department of Electrical and Computer Engineering
University of Massachusetts Lowell
Yan_Luo@uml.edu
http://acanets.uml.edu/~yluo/
Learning with Purpose
Outline
NSF IRNC AMIS Project
NSF CICI STREAMS Project
HeteroSpark Project
Wish List for Future Research Cyberinfrastructure
Learning with Purpose
Yan Luo, FIRE-­‐GENI, Washington DC, Sept 17-­‐18, 2015
2
IRNC AMIS at a Glance
UMass Lowell
FLowell 10G
SDN Science
Network
UMass
Boston
UMass Net
10G
10G pending
MGHPCC
KyRON
Peering
10G state
network
I2
Cincinnati,OH
I2 AL3S 100G
(north) est.
Q3'15
CUDIMexico
I2
Level 3
San
Antonio,
TX
El
Paso, TX
I2
Huston, TX
10G
Univ. of Texas
El Paso
Internet2
AMPATH / FIU
I2 ION
10G
Nox
Pronto 3295
MCT2
Lousville
KY
I2 AL3S
100G(west)
est. Q4'14
control
switch
Univ. of Kentucky
AL2S
Optic
Tap
GENI
Brocade
MLX-4
MCT1
7609
FLR
AMIS
• 40+Gbps flow-granularity
network
measurement
control
measurement instrument
measurement
link
• Software defined measurement
• Preserving privacy of network flow info
• In-depth flow analytics
Learning with Purpose
Optic
Tap
10G
FIU
Flowsurge
Yan Luo, FIRE-­‐GENI, Washington DC, Sept 17-­‐18, 2015
OpenFlow
Brazil
Domain
3
Challenges in Measurement Data Analytics
Diversity of analytics algorithms
• Privacy preserving algorithms: differential privacy,
searchable encryption, etc.
• reports of network activities: traffic matrix, flowlevel, burst-level
• Predictive: e.g. congestion event prediction
Variations on Traffic Load
• variations on compute loads
• Variations of network I/O demands
• Streaming requirements
Learning with Purpose
Yan Luo, FIRE-­‐GENI, Washington DC, Sept 17-­‐18, 2015
4
Measurement Data Management and
Processing in the Cloud
• A centralized
Operational Data
Management System to
manage distributed
AMAs
–
–
–
–
–
configuration
collection
storage
Processing
Reporting
Learning with Purpose
Yan Luo, FIRE-­‐GENI, Washington DC, Sept 17-­‐18, 2015
5
STREAMS : Secure Transport and REsearch
Architecture for Monitoring Stroke Recovery
Stroke and patient care
• stroke is one of the leading cause of death
• Stroke recovery is very time consuming and
• post-stroke monitoring of physiological status,
dietary intake, and rehabilitation progress for each
individual patient provides optimal patient recovery
outcomes.
Stroke patient monitoring
• Pervasive with multiple multi-modal sensors
• Smart watch, shimmer motes, google class, MS kinect
• Requires real-time data analytics for immediate
feedback
Learning with Purpose
Yan Luo, FIRE-­‐GENI, Washington DC, Sept 17-­‐18, 2015
6
Network Architecture Overview of STREAMS
Chameleon
CloudLab
Patient
home
Chattanooga
TN
Salt Lake City
Chicago
AL2S
GRE
Atlanta
UMass
Medical
School UMass
Lowell
Internet2
Boston
VPN
Layer 2
Pleiades Foundation
Erlanger Southeast Regional Stroke Center Learning with Purpose
Yan Luo, FIRE-­‐GENI, Washington DC, Sept 17-­‐18, 2015
MGHPCC
7
Security-ware VM Provisioning in the Cloud
Learning with Purpose
Yan Luo, FIRE-­‐GENI, Washington DC, Sept 17-­‐18, 2015
8
HeteroSpark: Heterogeneous Architecture for
Data-Compute-Intensive Applications
CPU Worker JVM
HeteroSpark
Call
Remote
Method
Serialized
Data
Deserialized
Data
RMI
Client
GPU Worker JVM
Input
Output
RMI
Server
Deserialized
Data
Remote
Method
JNI
ImplemenSerialized
tation
Data
Connected locally or
over the network
Learning with Purpose
Yan Luo, FIRE-­‐GENI, Washington DC, Sept 17-­‐18, 2015
9
Existing
Libs
*.so
Diverse Research Calls for Software Defined
Heterogeneous Cyberinfrastructure
Heterogeneous Architecture
• CPU/GPU/FPGA, x86/arm, Ethernet/infiniband
Software Defined Networking
• OpenFlow, ODL
Different Levels of Resource Abstraction
• VM, vCPU, vGPU
• Exclusive access to or reconfig physical resources
Sustainable Cost Model
• Encourage resource contribution
Learning with Purpose
Yan Luo, FIRE-­‐GENI, Washington DC, Sept 17-­‐18, 2015
10
Thank You !
Q&A
Learning with Purpose
Yan Luo, FIRE-­‐GENI, Washington DC, Sept 17-­‐18, 2015
11