Grid simulation (AliEn) Outline GRID simulation Simulation tool - - PowerPoint PPT Presentation
Grid simulation (AliEn) Outline GRID simulation Simulation tool - - PowerPoint PPT Presentation
Eugen Mudni Technical university Split -FESB Grid simulation (AliEn) Outline GRID simulation Simulation tool Ptolemy (Berkeley) Simulation architecture Simulation software environment Some results 2 GRID simulation
2
Outline
GRID simulation Simulation tool – Ptolemy (Berkeley) Simulation architecture Simulation software environment Some results
3
GRID simulation
- Why GRID simulation:
- 1. Validating grid architecture and scalability
- 2. System resources planning (storage
size,network throughput, processing power)
- 3. Simulating different job placement and file
transfer scheduling scenarios
- 4. GRID services tuning
- 5. Designing and testing of user strategies (how
to efficiently use the system)
4
Modeling and simulation
Top-down or Bottom-up ? Simulated system is very large and
complex (and still evolving)
Appropriately high degree of abstraction Focusing on the Grid key components
Storage elements Network Computing elements
Method : Discrete event simulation (DES)
5
Ptolemy Classic (Berkley)
Pros:
Fast (C++), stable Good DE library (component and messaging
system)
Free , source code available
Cons:
Complicated (CPP,TCL,PL(Ptolemy language),…) Resulting simulation needs Ptolemy installation Ask other Ptolemy users
DES tool - Ptolemy
6
C++
- bject definitions
- sim. control
- bject construction
connecting C++
Ptolemy customization
TCL Script
(simulation control,
- bject construction,
connecting)
PL (object definitions) C++
preprocessing compiling
- bj
Ptolemy libraries
- bj
- bj
PL (object definitions) PL (component definition) DE engine C++ source
7
AliEn DES model & DE components
server
CE CE CE Network file transfer simulation Storage Storage FTD FTD Message dispatcher
Job generation (simplified JDL)
FILE catalog
8
Network file transfer calculation component
Network as a set of links shared by changeable number
- f data streams
SE1 SE2 SE3 SE4
100 MB/s 30MB/s 50 MB/s 30 MB/s 1 GB/s 1 GB/s 100 MB/s
LAN1 LAN2 LAN3
links with capacity ( C1,..CL ) N streams , every stream has a predefined transfer route ri ⊂ {1,…,L} Every stream has equal priority Stream bandwidth allocation must conform to:
( )
i
i l r l
x C φ
∋
≤
∑
9
Network
Storage elements
Level 2 tape units
RW-access times, throughputs
Level 1 (HDD) Cache (CLC)
MB/s MB/s MB/s
CE
MB/s
MSS
10
Software environment
AliEnSim simulation application (Ptolemy C++)
MySQL database File catalog (directories) Simulation results system state dump
Java GUI
(system configuration editor, job description, visualization)
System configuration
LDAP
Simulation results & System state dump
X M L f i l e
Results GRID system configuration File catalog
- perations
11
¸Java GUI – grid configuration
12
Java GUI – job description
13
Java GUI – grid visualization(control)
14
Results example (jobs running)
15
Some performance testing results
Simulation of Alice PDCIII-Phase I:
104000 jobs execution , output is 3.8mil files (36 per
job)
15 sites (total 2100 CPU) central storage at CERN::Castor(MSS) &
CERN::scratch
Simulated real time: 10days Simulation time (P4-2GHz):
- ~1h
Simulation time
Other 20% Network 20% MySQL(file catalog) 60%
16
Future work ?
Test and enhance component models
Identify and use different system measuring
Network – throughput SE (MSS) : throughput, R/W access times
Optimize,optimize,…
Faster real system (more computing elements) <->