Transcript Satoshi_panel_ccgrid

"Towards Petascale Grids as a
Foundation of E-Science"
Satoshi Matsuoka
Tokyo Institute of Technology /
The NAREGI Project, National Institute of Informatics
Oct. 1, 2007 EGEE07 Presentation @ Budapest, Hungary
Vision of Grid Infrastructure in the
past…
OR
Bunch of networked
PCs virtualized to be
a Supercomputer
Very divergent &
distributed
supercomputers,
storage, etc. tied
together &
“virtualized”
The “dream” is for the infrastructure to behave as a
virtual supercomputing environment with an ideal
programming model for many applications
But this is not meant to be
Don Quixote or wrong tree dog
bark picture here
TSUBAME: the first 100 Teraflops
Supercomputer for Grids 2006-2010
Voltaire ISR9288 Infiniband
10Gbps x2 ~1310+50 Ports
~13.5Terabits/s
(3Tbits bisection)
10Gbps+External NW
Sun Galaxy 4 (Opteron Dual
core 8-socket)
Supercomputer in
10480core/655Nodes
th
Asia” 29
32-128GB
[email protected]
21.4TeraBytes
50.4TeraFlops
Now 103
OS Linux (SuSE 9, 10)
Unified IB TeraFlops Peak
NAREGI Grid MW
NEC SX-8i
(for porting)
Sun Blade
Integer Workload
Accelerator
(90 nodes, 720 CPU
network
“Fastest
as of Oct. 31st!
500GB
48disks 500GB
500GB
48disks
48disks
Storage
1.5PB 1.0 Petabyte (Sun “Thumper”)
0.1Petabyte (NEC iStore)
Lustre FS, NFS, CIF, WebDAV (over IP)
60GB/s 50GB/s aggregate I/O BW
ClearSpeed CSX600
SIMD accelerator
360 648 boards,
35
52.2TeraFlops
TSUBAME Job Statistics
Dec. 2006-Aug.2007 (#Jobs)
TSUBAME Jobs
700000
600000
500000
# Jobs
• 797,886 Jobs (~3270
daily)
• 597,438 serial jobs
(74.8%)
• 121,108 <=8p jobs
(15.2%)
90%
• 129,398 ISV Application
Jobs (16.2%)
• However, >32p jobs
account for 2/3 of
cumulative CPU usage
400000
系列1
300000
200000
100000
0
=1p
<=8p <=16p <=32p <=64p <=128p >128p
# Processors / Job
Coexistence of ease-of-use in both
- short duration parameter survey
- large scale MPI
Fits the TSUBAME design well
In the Supercomputing Landscape,
Petaflops class is already here… in early 2008
Other Petaflops 2008/2009
- LANL/IBM “Roadrunner”
- JICS/Cray(?) (NSF Track 2)
- ORNL/Cray
- ANL/IBM BG/P
- EU Machines (Julich…)
…
2008 LLNL/IBM “BlueGene/P”
~300,000 PPC Cores, ~1PFlops
~72 racks, ~400m2 floorspace
~3MW Power, copper cabling
2008Q1 TACC/Sun “Ranger”
~52,600 “Barcelona” Opteron
CPU Cores, ~500TFlops
~100 racks, ~300m2 floorspace
2.4MW Power, 1.4km IB cx4
copper cabling
2 Petabytes HDD
> 10 Petaflops
> million cores
> 10s Petabytes
planned for 2011-2012
in the US, Japan, (EU),
(other APAC)
In fact we can build one now (!)
• @Tokyo---One of the Largest IDC in the World
(in Tokyo...)
• Can fit a 10PF here easy (> 20 Rangers)
• On top of a 55KV/6GW Substation
• 150m diameter (small baseball stadium)
• 140,000 m2 IDC floorspace
• 70+70 MW power
• Size of entire Google(?) (~million LP nodes)
• Source of “Cloud” infrastructure
Gilder’s Law – Will make thin-client
accessibility to servers essentially “free”
Performance per Dollar Spent
Optical Fiber
(bits per second)
(Doubling time 9 Months)
Silicon Computer Chips
(Number of Transistors)
(Doubling time 18 Months)
0
1
Data Storage
(bits per square inch)
(Doubling time 12 Months)
2
3
Number of Years
(Original
Scientific American, January 2001
4
5
slide courtesy Phil Papadopoulos @ SDSC)
DOE SC Applications Overview
(following slides courtesy John Shalf @ LBL NERSC)
NAME
Discipline
Problem/Method
Structure
MADCAP
Cosmology
CMB Analysis
Dense Matrix
FVCAM
Climate Modeling
AGCM
3D Grid
CACTUS
Astrophysics
General Relativity
3D Grid
LBMHD
Plasma Physics
MHD
2D/3D Lattice
GTC
Magnetic Fusion
Vlasov-Poisson
Particle in Cell
PARATEC
Material Science
DFT
Fourier/Grid
SuperLU
Multi-Discipline
LU Factorization
Sparse Matrix
PMEMD
Life Sciences
Molecular Dynamics
Particle
Latency Bound vs. Bandwidth Bound?
• How large does a message have to be in order to
saturate a dedicated circuit on the interconnect?
– N1/2 from the early days of vector computing
– Bandwidth Delay Product in TCP
System
Technology
MPI Latency
Peak
Bandwidth
Bandwidth
Delay Product
SGI Altix
Numalink-4
1.1us
1.9GB/s
2KB
Cray X1
Cray Custom
7.3us
6.3GB/s
46KB
NEC ES
NEC Custom
5.6us
1.5GB/s
8.4KB
Myrinet Cluster
Myrinet 2000
5.7us
500MB/s
2.8KB
Cray XD1
RapidArray/IB4x
1.7us
2GB/s
3.4KB
• Bandwidth Bound if msg size > Bandwidth*Delay
• Latency Bound if msg size < Bandwidth*Delay
– Except if pipelined (unlikely with MPI due to overhead)
– Cannot pipeline MPI collectives (but can in Titanium)
(Original slide courtesy John Shalf @ LBL)
Diagram of Message Size Distribution
Function （MADBench-P2P)
60% of messages > 1MB
 BW Dominant, Could be
executed on WAN
(Original slide courtesy John Shalf @ LBL)
Message Size Distributions
(SuperLU-PTP)
> 95% of messages < 1KByte
 Low latency, tightly coupled LAN
(Original slide courtesy John Shalf @ LBL)
Collective Buffer Sizes
- demise of metacomputing -
95% Latency Bound!!!
=> For metacomputing,
Desktop and small
cluster grids pretty
much hopeless except
parameter sweep apps
(Original slide courtesy John Shalf @ LBL)
So what does this tell us?
• A “grid” programming model for parallelizing a single app
is not worthwhile
– Either simple parameter sweep / workflow, or will not work
– We will have enough problems programming a single system with
millions of threads (e.g., Jack’s keynote)
• Grid programming at “diplomacy” level
– Must look at multiple applications, and how they compete /
coordinate
• The apps execution environment should be virtualized --- grid being
transparent to applications
• Zillions of apps in the overall infrastructure, competing for
resources
• Hundreds to thousands of application components that coordinate
(workflow, coupled multi-physics interactions, etc.)
– NAREGI focuses on these scenarios
Use case in NAREGI: RISM-FMO
Coupled Simulation
Electronic structure of Nano-scale molecules in solvent is
calculated self-consistent by exchanging solvent charge
distribution and partial charge of solute molecules.
RISM
Solvent distribution
Mediator
GridMPI
FMO
Electronic structure
Mediator
Solvent charge distribution
is transformed from
regular to irregular meshes
Suitable for SMP
Suitable for Cluster
Mulliken charge is
transferred for partial
charge of solute
molecules
*Original RISM and FMO codes are developed by Institute of Molecular Science and
National Institute of Advanced Industrial Science and Technology, respectively.
Registration & Deployment of
Applications
Application Summary
Program Source Files
Input Files
Resource Requirements
etc.




Application
Developer
PSE Server
ACS
(Application Contents Service)
①Register Application
②Select Compiling Host
⑦Register
Deployment Info.
⑤Select Deployment Host
Application sharing in
Research communities
Compiling
OK!
Test Run
OK!
Server#1
⑥Deploy
Test Run
Server#2
NG!
Test Run
Server#3
Resource
Info.
Information Service
OK!
Description of Workflow and Job
Submission Requirements
http(s)
tomcat
Web server(apache)
applet
Data icon
/gfarm/..
Program icon
Appli-A
JSDL
Workflow
Servlet
JSDL
Global file
information
Application Information
PSE
NAREGI
JM I/F module
BPEL
<invoke name=EPS-jobA>
↓ JSDL -A
<invoke name=BES-jobA>
↓ JSDL -A
…………………..
Appli-B
DataGrid
Wokflow Description
By NAREGI-WFML
Information
Service
GridFTP
(Stdout
Stderr)
BPEL+JSDL
Super
Scheduler
Server
Reservation Based CoAllocation
• Co-allocation for heterogeneous architectures and
applications
• Used for advanced science applications, huge MPI jobs,
realtime visualization
on grid, etc...
Workflow
Abstract
JSDL
Client
Reservation based
Co-Allocation
Super
Scheduler
Resource
Query
DAI
Reservation, Submission,
Query, Control…
Concrete
JSDL
GridVM
Distributed
Information Service
Concrete
JSDL
CIM
Resource
Info.
Accounting
GridVM
UR/RUS
Computing Resource
Computing Resource
Communication Libraries and Tools
1.
Modules
2.
Features
GridMPI:
MPI-1 and 2 compliant grid ready MPI library
GridRPC:OGF/GridRPC compliant GridRPC library
Mediator:
Communication tool for heterogeneous applications
SBC:
Storage based communication tool
GridMPI
MPI for a collection of geographically distributed resources
High performance optimized for high bandwidth network
GridMPI
Task parallel simple seamless programming
Mediator
Communication library for heterogeneous applications
Data format conversion
SBC
Storage based communication for heterogeneous applications
3.
Supporting Standards
MPI-1 and 2
OGF/GridRPC
Grid Ready Programming
Libraries
• Standards compliant GridMPI and GridRPC
GridMPI
Data Parallel
MPI Compatibility
GridRPC (Ninf-G2)
Task Parallel, Simple
Seamless programming
RPC
100000 CPU
RPC
100-500 CPU
Communication Tools for CoAllocation Jobs
• Mediator
Application-1
Mediator
Mediator
Data Format
Conversion
Data Format
Conversion
GridMPI (
Application-2
)
• SBC (Storage Based Communication)
Application-3
SBC library
SBC library
SBC protocol (
)
Application-2
Compete Scenario: MPI / VM Migration on
Grid (our ABARIS FT-MPI)
Cluster A (fast CPU, slow networks)
Resource
Manager
MPI
MPI
MPI
VM
Host
VM
Host
VM
Host
Host
App A
(High BW)
Resource manager, aware
of individual application
characteristics
MPI Comm Log
Redistribution
Host
Host
VM Job Migration
Power Optimization
App B（CPU)
App A (High Bandwidth)
MPI
MPI
MPI
MPI
MPI
MPI
VM
Host
VM
Host
VM
Host
VM
Host
VM
Host
VM
Host
Cluster B (high bandwidth, large memory)