emerging hw architectures
TRANSCRIPT
HPCMP
HPCMP
Emerging HW ArchitecturesDoD C2 SW requirements lag HW power
Polymorphous Computing Architectures
(1999-2015)
Cognitive Computing(2003-2023)
Adaptive/Reconfigurable Computing Systems
(1991-2010)
High Productivity Computing Systems
(2001-2012)
Power Aware Computing(1999-2006)
Data Intensive Computing(1998-2006)
Peta-Scale Systems(2005-2018)
Status
• 8 Codes Tested and Integrated with Framework
• JBI Core Services Demonstrated
Impact
• Near Real-Time Access to Hyperspectral Imagery Data for Battlespace Awareness
• Imagery Products for Intelligence Analysts and Battlefield Decision Makers
Approach
• Leverage serial DoD hyperspectral imagery codes to rapidly process raw hyperspectral data to produce imagery products
• Client for JBI to provide rapid query, publish and subscribe with secure webserver application interface
Hyperspectral Imaging PortfolioObjective: To develop a scalable, portable, high performance computing software framework with 8
algorithms for rapidly accessing and processing hyperspectral data.
Pictorial description ofproject
Hyperspectral Cubes
DREN
HuinaluClusterServer
HeteroHPC
Server
Rack 2 KerberizedJBI Pub/Sub Broker
ARL MSRCBrainerdIBM P3
Rack 3KerberizedFramework
RAID(Oracle
Database)
RomeUsers
OutsideUsers
CoyoteHPC
Server
HuinaluMHPCC
HIE Test Framework Grid
Rack 1 KerberizedJBI Repository
1
2
3
N
Parallelized HIEApplications
HIE Server
JOB Submittal Result Retrieval JBI QueryCubes
JBI HPC StatusSubscriber
Parallel Hyperspectral Application ServerHPC JobLauncher
Results Manager
Cube Publisher &Subscriber
StatusPublisher
khttps
kftpXML
kftpNITF
NFS Mounted Shared Disk
JBI
NITFCube on
Local Disk
FrameworkApp Interface
Program
MPINode
0
Disk I/O
CUBE
Cho
pper
Scat
tere
rG
athe
rer
Gen
erat
orNITF OutputApp Specific
App Specific
Disk I/O
NITF
Pars
er
NITF Viewers
HPC/JBI/Server Interplay
NSF TeraGrid
Distributed HPC’sAccessed by authorized users anywhere on the DREN and InternetInteractive and time critical problems
DPIM TestbedDPIM Testbed
LegendRemote UsersNetworked HPC’s
Defense Research and Engineering Networks
AFRLCoyoteHHPC
ARLPowell
ASCMach 2
SSCSDSeahawk
MHPCCKoaCluster
DPIM-DREN Interactive Network
HPC1 HPC2
HPC4HPC3
coyote
First Client Connection
List of Peers
HHPC
Client Connection
Client Connection
seafarer
powell
HPCs are all interconnected
Client gets list of HPCs from first connection
Client maintains up to three connections
Dynamic load balancing between HPCs based on I/O type and activity
The client receives copies in triplicate of each publication
Joint Battlespace Infosphere
• Current JBI:– Provide low latency and high capacity of the core
services of a distributed, fault tolerant information management system
• Goal: – Use distributed interactive HPC resources on DREN– Fault tolerance– Scalability
Scalability
•300 Predicates•600 Clauses•0.3% hit ratio (1/300)•4-6 Publishers were used•2 KB IOs
•300 subscribers•Predicate ensures that each publication is delivered to exactly one subscriber
Scalability
0100020003000400050006000700080009000
0 5 10 15 20
Number of Brokers
Max
Agg
rega
te R
ecei
ve R
ate
(io/s
) Max RatesOptimum
HPCMP
HPCMP
HPCMP
HPCMP
HPCMP
HPCMP
HPCMP
HPCMP
HPCMP
HPCMP
HPCMP
HPCMP
HPCMP
HPCMP