advanced rti system (artìs) - unibo.itgdangelo.web.cs.unibo.it/pool/ricerca/artis-static...
TRANSCRIPT
Gabriele D’AngeloMichele BracutoLuciano Bononi
{gda, bracuto, bononi}@cs.unibo.it
http://www.cs.unibo.it/~gdangelohttp://www.cs.unibo.it/~bracutohttp://www.cs.unibo.it/~bononi
Dipartimento di Scienze dell’InformazioneUniversità degli Studi di Bologna
Advanced RTI System (ARTìS)A new middleware for parallel
and distributed simulation
2ARTìS 2005© 2005 Gabriele D’Angelo
Presentation outline
Simulation of large scale and complex models
The ARTìS software architecture
Simulation of massive, complex and dynamic systems
High Performance Computing (HPC)
Gaia framework: entities migration
Examples: Ad hoc and Sensors network models
Concurrent Replication of PADS (CR-PADS)
Further optimizations for PADS
Conclusions and future work
3ARTìS 2005© 2005 Gabriele D’Angelo
Simulation of large scale and complex models
Simulation models currently of interest may include a potentially
huge number of simulated objects
Large scale and complex simulation models may be unpractical to
simulate on a single-processor execution unit: huge memory
requirements, large amount of time required to complete the
simulation runs
The memory bottleneck reduction, scalability and speed-up can
be achieved by using parallel/distributed models and execution
architectures
Goal: to increase the simulation speed, reduce the Wall Clock
Time (WCT) required to complete the simulation runs
4ARTìS 2005© 2005 Gabriele D’Angelo
The ARTìS software architecture
Advanced RTI System (ARTìS), parallel and distributed simulation
middleware:
Model components’ heterogeneity, distribution and reuse
Adaptive evaluation of the communication bottlenecks and
dynamic adaptation of the inter-process communication layer
Generic Adaptive Interaction Architecture (GAIA): model
components’ migration mechanism to support load balancing and
data distribution management (DDM) overhead reduction
Support for High Performance Computing clusters (HPC):
scalability evaluation of the middleware
Concurrent Replication of Parallel and Distributed Simulation (CR-
PADS) and cloning
5ARTìS 2005© 2005 Gabriele D’Angelo
ARTìS: logical architecture
SharedMemory TCP/IP R-UDP Multicast
DDM Time Man.
FederationMan.
DeclarationMan.
ObjectMan.
Own. Man.
GamingAPI
HLAIEEE 1516
User Simulation Level
Unibo API
TCP/IP C/C++ Real-TimeIntrospection
Logging
PerformanceCommunicationLayer
RTI Core
RTI Interfaces(API)
Cloning & Replication
MPI
JAVA
GAIA
6ARTìS 2005© 2005 Gabriele D’Angelo
HOST C
HOST B
HOST A
Communication architecture
SIMA
SHM TCP/IP
LP #3
SHM TCP/IP
LP #1
SHM TCP/IP
LP #0
TCP/IP
LP #4
SHM TCP/IP
LP #2
7ARTìS 2005© 2005 Gabriele D’Angelo
Performance evaluation: Ad-Hoc wireless network model
Simulated model:
A set of Simulated Mobile Hosts (SMHs)
Mobility model:
Random Mobility Motion model (RMM)
uncorrelated SMHs’ mobility
Traffic model:
ping messages (CBR) by every SMH to all neighbors within
the wireless communication range (250 m)
Propagation model
open space (neighbor-SMHs within detection range)
8ARTìS 2005© 2005 Gabriele D’Angelo
Ad-Hoc network model characterization
Computation and communication issues:
The computation required for each SMH per time-step is in the
order of O(#SMH^2): to determine the neighbor set
The communication required among SMHs is in the order of
O(K*#SMH) per time-step, with K defined as a constant value
based on SMHs density (assumed as constant)
9ARTìS 2005© 2005 Gabriele D’Angelo
Scalability evaluation: High Performance Computing
IBM CLX/1024 – IBM Linux cluster 1350 - CINECA
512 2-way nodes (IBM X335)
768 Xeon Pentium IV, 3.06 GHz
256 Xeon Pentium IV EM64T (Nocona), 3.00 GHz
2 GB of RAM for each node
Global peak performance: 6.1 TFlops
All the nodes are interconnected by a low latency Myrinet network,
maximum bandwidth between each pair of nodes: 256 MB/s
10ARTìS 2005© 2005 Gabriele D’Angelo
ARTìS on High Performance Computing (HPC)
ARTìS on HPC Cluster
0
500
1000
1500
2000
2500
3000
3500
4000
4500
0 16 32 48 64 80 96 112 128
LPs (CPUs)
Wal
l Clo
ck T
ime
(s)
ARTìS Speedup on HPC Cluster
0
1
2
3
4
5
6
7
8
9
10
0 16 32 48 64 80 96 112 128
LPs (CPUs)
Spee
dup
1 million of Simulated Mobile Hosts (1 Timestep of simulated time)
11ARTìS 2005© 2005 Gabriele D’Angelo
“Open broadcast” nature of the wireless transmissions
“space-locality” of causality between neighbor-hosts
neighbor-hosts should be notified about transmission events anyway,
e.g. to model interference, detection, MAC, etc.
Wireless devices can be mobile
the set of neighbor-hosts change as simulated time elapses
Communication between hosts is “session-based”
determines a “time-locality” effect
the set of neighbor-hosts is interested by transmission events, for a
significant time-window
The group of model entities in the shared causal-domain can be
highly dynamic:
high degree of communication is required to maintain full synchronization
Mobile and Wireless Networks’ model characteristics
12ARTìS 2005© 2005 Gabriele D’Angelo
GAIA framework: entities migration
Wireless ad hoc network scenario: (evaluating migration of SMH x)
X’s “transmission-event” must be notified to the 4 model entities executed over B
After X’s migration, X’s “transmission-event” must be notified to one model entity executed over A
A B
network delay
Physical Execution Units for the simulation
model entities of SMHsexecuted over PEUs A and B
Host X executed over PEU A“transmission-event”
A B
network delay
Physical Execution Units for the simulation
model entities of SMHsexecuted over PEUs A and B Same scenario after migration of X from A to B
X X
X
SMH = Simulated Mobile HostPEU = Physical Execution Unit
13ARTìS 2005© 2005 Gabriele D’Angelo
Example: Ad-Hoc network model implementation
Modeling issues:
A set of Simulated Mobile Hosts (SMHs)
Mobility model:
Random Mobility Motion model (RMM)
Fast- and Slow-RMM (100, 25, 10 m/s)
uncorrelated SMHs’ mobility (worst case)
Traffic model:
ping messages (CBR) by every SMH to all neighbors within
the wireless communication range (250 m)
low local computation model (worst case)
Propagation model
open space (neighbor-SMHs within detection range)
14ARTìS 2005© 2005 Gabriele D’Angelo
Ad hoc network: migration mechanism “off” and “on”
Migration “OFF” Migration “ON”
17ARTìS 2005© 2005 Gabriele D’Angelo
Performance analysis: speed-up
0,0
0,5
1,0
1,5
2,0
2,5
3,0
Spe
ed-u
p
M=1, N=1 M=1, N=3 M=2, N=2 M=2, N=4 M=3, N=3 M=3, N=6
Ad Hoc (25 m/s): Speed-up
Migration OFF Migration ONM = # PEU, N = # LP
18ARTìS 2005© 2005 Gabriele D’Angelo
Example: sensor network model implementation
Design of a new energy aware Medium Access Control protocol:
Up to 40.000 sensors, limited battery resources
Sensors are static (no mobility model)
Every sensor implements the MAC protocol, no centralization
4 different sensor states (active, power saving, listening, died)
Every sensor implements a “pressure variation” detector and
sends broadcast alerts flooding toward a set of detection points
The energy aware MAC protocol increases the network lifetime
managing the sensors’ state (dinamically switching to power-save
state the sensors within areas covered by other active sensors)
19ARTìS 2005© 2005 Gabriele D’Angelo
Sensor network simulation example
1000 sensors (for this example),Limited battery resources
4 different states:• red = awake• green = power saving• white = listening• black = died
Pink circle = active transmitting range
Blue circle = alert-message transmission
GOAL: extend the network lifetime maintaining network connectivity
21ARTìS 2005© 2005 Gabriele D’Angelo
Conclusion and Future work
ARTìS is a scalable, optimized parallel and distributed simulation
middleware, used to simulate dynamic complex systems
Careful and enhanced evaluation of the proposed optimizations
Further improve the ARTìS middleware
Data structures optimization (event-management)
GAIA: detailed load balancing and communication heuristics
IEEE 1516 full compatibility
New models (es. Detailed MAC protocols)
Migration based middleware for Internet games
Gabriele D’AngeloMichele Bracuto
{gda, bracuto}@cs.unibo.it
http://www.cs.unibo.it/~gdangelohttp://www.cs.unibo.it//~bracuto
Dipartimento di Scienze dell’InformazioneUniversità degli Studi di Bologna
Advanced RTI System (ARTìS)A new middleware for parallel
and distributed simulation