Transcript Programming High Performance Applications using Components

Programming High Performance
Applications using Components
Thierry Priol
PARIS research group
IRISA/INRIA
http://www.irisa.fr/paris
A joint work with A. Denis, C. Perez and A. Ribes
Supported by the French ACI GRID initiative
Outline







High-Performance applications and code coupling
The CORBA Component Model
CCM in the context of HPC
GridCCM: Encapsulation of SPMD parallel codes into components
Runtime support for GridCCM components
Conclusions & future works
Some references
High Performance applications
Not anymore a single parallel application but several of them


High performance applications are more and more complex…
thanks to the increasing in performance of off-the-shelves
hardware
Several codes coupled together involved in the simulation of
complex phenomena



Virtual plane
Plane
wave
Scattering of 1 on 2
Fluid-fluid, fluid-structure, structure-thermo, fluid-acoustic-vibration
Even more complex if you consider a parameterized study for
optimal design
Some examples

e-Science

Industry



Scattering of 2 on 1
Object 1
Object 2
Single-physic multiple-object
Weather forecast: Sea-Ice-Ocean-Atmosphere-Biosphere
Aircraft: CFD-Structural Mechanics, Electromagnetism
Satellites: Optics-Thermal-Dynamics-Structural Mechanism
Structural
Mechanics
Optics
Quic kT ime™ et un déc ompres s eur T IFF (non compress é) sont requis pour vi si onner c ette image.
Thermal
Dynamics
Multiple-physics single-object
Electromagnetic coupling
EuroPVM/MPI 03 - Sept., 2003
Ocean-atmosphere coupling
Programming High Performance Applications using Components
2
The current practice…

Coupling is achieved through the use of
specific code coupling tools

Not just a matter of communication !



Interpolation, time management, …
Examples: MpCCI, OASIS, PAWS,
CALCIUM, ISAS, …
MPI
code
#1
MPI
code
#2
MPI
code
#3
Code coupler (MpCCI, OASIS …)
MPI implementation
Limitations of existing code coupling tools






Originally targeted to parallel machines with
some on-going works to target Grid
infrastructure
Static coupling (at compile time): not “plug
and play”
Ad-hoc communication layers (MPI,
sockets, shared memory segments, …)
Lack of explicit coupling interfaces
Lack of standardization
Lack of interoperability
…
MPI
code
#n
SAN
cluster
Code A
Code B
MpCCI
MpCCI
MPI
The MpCCI coupling library
EuroPVM/MPI 03 - Sept., 2003
Programming High Performance Applications using Components
3







Codes are encapsulated into components
Components have public interfaces
Components can be coupled together or through
a framework (code coupler)
Components are reusable (with other
frameworks)
Application design is simpler through composition
(but component models are often complex to
master !)
Some examples of component models


EuroPVM/MPI 03 - Sept., 2003
In
Code B
Code
interface
Code
interface
In
Out
Out
In
In
In
In
In
Out
Code coupler
Component-based framework
Out
Out
CCA, ICENI
Standard component models

Code A
Out
HPC component models

Out
Code A
Component programming is well suited for
code coupling
In

“A component is a unit of independent
deployment”
“Well separated from its environment and from
other components”
Code B

Code
interface
Component definition by C. Szyperski
Out

Code
interface
Another approach for code coupling
EJB, DCOM/.NET, OMG CCM
Programming High Performance Applications using Components
Out
Out
In
In
Code
interface
Code
interface
Code C
Code D
4
Distributed components: OMG CCM
Component interface







An architecture for defining components
and their interaction
Interaction implemented through
input/output interfaces
Synchronous and asynchronous
communication
A packaging technology for deploying
binary multi-lingual executables
A runtime environment based on the
notion of containers (lifecycle, security,
transaction, persistent, events)
Multi-languages, multi-OS, multi-ORBs,
multi-vendors, …
OFFERED
A distributed component-oriented model
Facets
My
Component
Event
sinks
Receptacles
Event
sources
REQUIRED

Attributes
Include a deployment model

Could be deployed and run on several
distributed nodes simultaneously
EuroPVM/MPI 03 - Sept., 2003
Component-based application
Programming High Performance Applications using Components
5
From CORBA 2.x to …





Open standard for distributed object
computing by the OMG
Software bus, object oriented
Remote invocation mechanism
Hardware, operating system and
programming language independence
Vendor independence (interoperability)
interface MatrixOperations {
const long SIZE = 100;
typedef double Vector[ SIZE
typedef double Matrix[ SIZE
void multiply( in Matrix A,
out Vector C
};
EuroPVM/MPI 03 - Sept., 2003
];
][ SIZE ];
in Vector B,
);
IDL file
Server
IDL
Compiler
Client
Obj. Impl.
Obj. Inv.
IDL Skel.
IDL Stub
POA
Object Request Broker (ORB)
GIOP
ESIOP
IIOP
Programming High Performance Applications using Components
6
… CORBA 3.0 and CCM
interface Mvop
{
void mvmult(some parameters);
};
interface Stat
{
Void sendt(some parameters);
}
component MatrixOp
{
attribute long size;
provides Mvop the_mvop;
uses Stat the_stat;
publishes StatInfo avail;
consume MustIgo go;
}
OFFERED
Component interface specified by the
OMG IDL 3.0
MvopHome
Component interface
Facet
REQUIRED

Receptacle
MatrixOp
Event
sink
Event
source
Attribute
Home MvopHome manages MatrixOp {};
EuroPVM/MPI 03 - Sept., 2003
Programming High Performance Applications using Components
7
CCM Runtime Environment &
Deployment
Container

CCM includes a runtime environment
called Container


Component
Callback Implementation
A container is a component’s only
outside contact
API
CCM defines a Packaging and
Deployment model

ORB
Persistency
Components are packaged into a selfdescriptive package (zip)

Internal
Interfaces
Transaction
One or more implementations (multibinary) + IDL of the component + XML
descriptors
Notification
Security
Install
Object
Install
Processor
Object
Processor

Packages can be assembled (zip)

A set of components + XML descriptor
ZIP
Deployment App
OMG IDL
Assembly File
Install
Object

Assemblies can be deployed
Processor

Through a deployment tool to be
defined… such as a Grid middleware…
Staging Area
[my_favorite_grid_middlware]run myapp.zip
Install
Object
Processor
EuroPVM/MPI 03 - Sept., 2003
Programming High Performance Applications using Components
Processor
8
CCM in the context of HPC

Encapsulation of parallel codes into software components


Parallelism should be hidden from the application designers as much as
possible when assembling components
Issues:



How much my parallel code has to be modified to fit the CCM model
What has to be exposed outside the component ?
Communication between components


Components should use the available networking technologies to let
component to communicate efficiently
Issues:



EuroPVM/MPI 03 - Sept., 2003
How to combine multiple communication middleware/runtime and to make
them to run seamlessly and efficiently ?
How to manage different networking technologies transparently ?
Can my two components communicate using Myrinet for one run and using
Ethernet for another run without any modification, recompilation,..?
Programming High Performance Applications using Components
9
Encapsulating SPMD codes into
CORBA Components
The obvious way : adopt a
master/Slave approach




Advantage


One SPMD process acts as the master
whereas the others act as slaves
The master drives the execution of the
slaves through message passing
Communication between the two SPMD
codes go through the master process
Could be used with any* CCM
implementation

Component B
MPI Master processes
MPI Master processes
SPMD
Proc.
SPMD
Proc.
MPI comm. layer
MPI comm. layer
SPMD
Proc.
SPMD
Proc.
SPMD
Proc.
SPMD
Proc.
SPMD
Proc.
SPMD
Proc.
SPMD
Proc.
SPMD
Proc.
…
MPI Slave processes
Drawbacks

Component A
ORB

Lack of scalability when communicating
through the ORB
Need modifications to the original MPI
code
Code A
…
MPI Slave processes
Code B
WAN
40 Gbit/s
Ethernet
Switch
100 Mb/s
Ethernet
Switch
100 Mb/s
100 Mb/s
100 Mb/s
* In theory … but practically…
EuroPVM/MPI 03 - Sept., 2003
Programming High Performance Applications using Components
10
Making CORBA Components
parallel-aware

Just to remind you :


Communication between
components is implemented
through a remote method
invocation (to a CORBA object)
Constraints




EuroPVM/MPI 03 - Sept., 2003
A parallel component with one
SPMD process = a standard
component
Communication between
components should use the ORB
to ensure interoperability
Little but preferably no modification
to the CCM specification
Scalable communication between
components by having several data
flows between components
What the application designer should see…
HPC
Component
A
HPC
Component
B
… and how it must be implemented !
// Comp. A
SPMD
Proc.
SPMD
Proc.
SPMD
Proc.
SPMD
Proc.
SPMD
Proc.
Programming High Performance Applications using Components
// Comp. B
SPMD
Proc.
SPMD
Proc.
SPMD
Proc.
SPMD
Proc.
SPMD
Proc.
11
Parallel Remote Method invocation
A set of processes calling…

Based on the concept of “multifunction” introduced by J-P.
Banâtre et al. (1986)


Parallel remote Method invocation



A set of processes collectively calls
another set of processes through a
multi-function (multi-RPC)
Notion of collection of objects
(parallel objects)
Multi-RMI
Parallel Component


EuroPVM/MPI 03 - Sept., 2003
A collection of identical
components
A collection of parallel interfaces
(provide/use)
X=1
Y=3
Call f(X,Y,Z)
A=Z*2
.
.
.
.
X=3
Y=5
Call f(X,Y,Z)
A=Z*2
.
.
.
.
F(A, B, C)
Begin
…
C=A+3*B
…
return
F(A, B, C)
Begin
…
C=A+3*B
…
return
…
…
X=7
Y=8
Call f(X,Y,Z)
A=Z*2
.
.
.
.
F(A, B, C)
Begin
…
C=A+3*B
…
return
… another set of processes
Parallel Component A Parallel Component B
Comp. A-0
Comp. A-1
A-0
SPMD
Comp. A-2
A-0
Proc.
SPMD
Comp. A-3
A-0
Proc.
SPMD
Comp. A-4
A-0
Proc.
SPMD
Proc.
SPMD
Proc.
Programming High Performance Applications using Components
Comp. B-0
Comp. B-1
SPMD
Comp. B-2
Proc.
SPMD
Comp. B-3
Proc.
SPMD
Comp. B-4
Proc.
SPMD
Proc.
SPMD
Proc.
12
Parallel interface

A parallel interface is mapped on a collection
of identical CORBA objects


Collection specification through IDL
extensions




Invocation to a collection of objects is
transparent to the client (either sequential or
parallel)
Size and topology of the collection
Data distribution
Reduction operations
Extended-IDL
compiler
Modification to the IDL compiler


Stub/skeleton code generation to handle parallel
objects
New data type for distributed array parameter



interface[*:2*n] MatrixOperations {
typedef double Vector[SIZE];
typedef double Matrix[SIZE][SIZE];
void multiply(in dist[BLOCK][*] Matrix A,
in Vector B,
out dist[BLOCK] Vector C);
void skal(in dist[BLOCK] Vector C,
out csum double skal);
};
Extension to CORBA sequence
Data distribution information
Impact on standard:



Does not require the modification of the ORB
Require extensions to IDL and naming service
Is not portable across CORBA implementations
EuroPVM/MPI 03 - Sept., 2003
Parallel CORBA Object
MPI comm. layer
Sequential
client
Object
Inv.
Stub
SPMD
Proc.
…
SPMD
Proc.
Skel.
Skel.
POA
POA
Object Request Broker (ORB)
Programming High Performance Applications using Components
13
Parallel invocation and data
distribution
Parallel CORBA Object
MPI comm. layer


Data management is carried out by the stub
and the skeleton
What the stub has to do:


generates requests according to the number
of objects belonging to the collection
According to the data distribution


Sequential
client
SPMD
Proc.
Object
Inv.
Stub
Scatter the data for the input parameters (in)
Gather the data for the output parameters (out)
…
Skel.
Skel.
POA
POA
Object Request Broker (ORB)
Skel.
Stub
...
Pco->foo(A)
...
Programming High Performance Applications using Components
Request1
Request2
Request3
SPMD
Proc.
foo( A )
foo( A )
foo( A )
MPI comm. layer
EuroPVM/MPI 03 - Sept., 2003
A
Request0
Request0
Object Request Broker (ORB)
Gather
interface[*] Operations {
typedef double Matrix[100][100];
void foo(inout dist[BLOCK][*] Matrix A);
};
SPMD
Proc.
foo( A )
Scatter
14
Parallel invocation and data
distribution
Parallel CORBA Object
MPI comm. layer

Different scenarios



Stubs synchronized themselves to elect those
No serialisation here
that call the server objects
But parallel connection


Parallel
client
M x 1 : the server is sequential
M x N : the server is parallel
Synchronisation through MPI
Data redistribution is performed by the stub
before calling the parallel CORBA object

Parallel
server
SPMD
Proc.
SPMD
Proc.
Stub
Stub
…
SPMD
Proc.
Stub
Object Request Broker (ORB)
POA
POA
POA
Skel.
Skel.
Skel.
SPMD
Proc.
SPMD
Proc.
Data redistribution library using MPI
…
SPMD
Proc.
MPI comm. layer
Parallel CORBA Object
Parallel Server
A[BLOCK][*]
A[*][BLOCK]
2) Parallel Invocation
ORB
Redistribution
for parameter A
Parallel Client
1) Redistribution
Parallel Server
Parallel Client
Data redistribution at client side
EuroPVM/MPI 03 - Sept., 2003
Programming High Performance Applications using Components
15
Lessons learnt…

Implementation dependant


Modification of the CORBA standard (Extended-IDL)



The OMG does not like such approach
The end-users too…
Data distribution specified in the IDL


Parallel CORBA object only available for MICO
Difficulties to add new data distribution schemes
Other approaches

PARDIS (Indiana University, Prof. D. Gannon)


Data Parallel CORBA (OMG)


EuroPVM/MPI 03 - Sept., 2003
Modification of the IDL language (distributed sequence, futures)
Require modifications to the ORB to support data parallel object
Data distribution specification included in the client/server code and given to
the ORB
Programming High Performance Applications using Components
16
Portable parallel CORBA object

No modification to the OMG standard




Parallelism specification through XML




Standard IDL
Standard ORB
Standard naming service
Operation: sequential, parallel
Argument: replicated, distributed
Return argument: reduction or not
Automatic generation of a new CORBA
object (Parallel Stub and Skeleton)


Renaming
Add new IDL interfaces
EuroPVM/MPI 03 - Sept., 2003
#include “Matrix.idl”
Interface IExample {
void
send_data(Matrix m);
};
Interface IExample
Method_name: send_data
Method_type: parallel
Return_type: none
Number_argument: 1
Type_Argument1:
Distributed
…
User IDL
Distribution
(XML)
Parallel IDL
compiler
Parallel IDL
Standard IDL
compiler
Parallel
CORBA
Stub & Skel.
Standard Stub
Standard Skel.
Programming High Performance Applications using Components
17
Portable parallel CORBA object

Behave like standard CORBA
objects


A parallel CORBA object is:



A collection of identical CORBA
object
A Manager object
MPI WORLD

MPI WORLD
Parallel User Code (MPI)
Interface1 Implementation
Interface1
Interface1
Parallel CORBA stub
Parallel CORBA skel
ManagerInterface1
ManagerInterface1
CORBA stub
Data distribution managed by the
Parallel CORBA stub and skeleton


I_var o = I::_narrow(...);
o->fun();
CORBA skeleton
CORBA ORB
At the client or the server side
Through the ORB
Interoperability with various CORBA
implementations
EuroPVM/MPI 03 - Sept., 2003
Programming High Performance Applications using Components
18
Back to CCM: GridCCM
interface Interfaces1
{
void example(in Matrix mat);
};
…
component A
{
provides Interfaces1 to_client;
uses
Interfaces2 to_server;
};
Parallel Component
Of type A
to_client
IDL
Component: A
Port: to_client
Name: Interfaces1.example
Type: Parallel
Argument1: *, bloc
…
EuroPVM/MPI 03 - Sept., 2003
to_server
Comp. A-0
Comp. A-1
A-0
SPMD
Comp. A-2
A-0
Proc.
SPMD
Comp. A-3
A-0
SPMD
Proc.
Comp. A-4
A-0
SPMD
Proc.
Proc.
SPMD
Proc.
XML
Programming High Performance Applications using Components
19
Runtime support for a grid-aware
component model

Main goal for such a runtime

Should support several communication runtime/middleware at the same time


Underlying networking technologies not exposed to the applications


Parallel runtime (MPI) & Distributed Middleware (CORBA) such as for GridCCM
Should be independent from the networking interfaces
Let’s take a simple example




MPI and CORBA using the same protocol/network (TCP/IP, Ethernet)
MPI within a GridCCM component, CORBA between GridCCM components
The two libraries are linked together with the component code, does it work ?
Extracted from a mailing list:
Message 8368 of 11692 | Previous |Next [ Up Thread ] Message Index
From: ------- -------- < -.-----@n... >
Date: Mon May 27, 2002 10:04 pm
Subject: [mico-devel] mico with mpich
I am trying to run a Mico program in parallel using mpich. When calling
CORBA::ORB_init(argc, argv) it seems to coredump. Does anyone have
experience in running mico and mpich at the same time?
EuroPVM/MPI 03 - Sept., 2003
Programming High Performance Applications using Components
20
PadicoTM: an open integration
framework

Provide a framework to combine various communication
middleware and runtimes

For parallel programming:



For distributed programming




Transparency vis à vis the applications
Get the maximum performance from the network!


RPC/RMI based middleware (DCE, CORBA, Java)
Middleware for discrete-event based simulation (HLA)
Avoid the NxM syndrome !
…
Std
Object

…
Std
Sim
HLA Framework
Middleware & runtimes
Grid
Services
MPI
CORBA
…
HLA
PadicoTM framework
Offer zero-copy mechanism to middleware/runtime
What are the difficulties ?

MPI
Sim.
CORBA Framework
To handle a large number of networking interfaces


Message based runtimes (MPI, PVM, …)
DSM-based runtimes (TreadMarks, …)
MPI code
Object
WAN
SAN
Provide a generic framework for parallel-oriented runtime (SPMDbased + fixed topology) and distributed-oriented middleware (MPMDbased + dynamic topology)
Multiplexing the networking interface


A single runtime/middleware use several networking interfaces at the same
time
Multiple runtime/middleware use simultaneously a single networking
interface
EuroPVM/MPI 03 - Sept., 2003
Homogeneous cluster
Programming High Performance Applications using Components
Supercomputer
Visualisation
21
PadicoTM architecture overview

Provide a set of personalities to
make easy the porting of existing
middleware


The Internal engine controls the
access to the network and
scheduling of threads


DSM
MPI
CORBA
Available on a large number of
networking technologies
Associated with Marcel
(multithreading)
EuroPVM/MPI 03 - Sept., 2003
Kaffe
JVM
CERTI
HLA
Personality Layer
Arbitration, multiplexing, selection,
…
Low level communication layer
based on Madeleine

Mpich
PadicoTM
Services
PadicoTM
Core
Internal engine
Madeleine
Portability across networks
Myrinet
SCI
Network
s

BSD Socket, AIO, Fast Message,
Madeleine, …
Mome
OmniORB
MICO
Orbacus
Orbix/E
TCP
Programming High Performance Applications using Components
Marcel
I/O aware multi-threading
Multithreading
22
PadicoTM implementation
SPMD process on each participating node

Implemented as a single process



To gain access to networks that cannot
be shared
Launched simultaneously on each
participating node
Middleware and user’s applications
are loaded dynamically at runtime

DSM
Applications
MPI
Circuit
CORBA
High Performance
Sockets
NetAccess
Network multiplexing
Polling/callback management
Madeleine
Portability across networks
Myrinet
JVM
VSock
Network Topology
Management
SCI
Network
s

Appear as modules
Modules can be binary shared objects
(“.so” libraries on Unix system) or Java
classes
User code
TCP
GK
TaskManager
Modules management
Centralized thread policy
Marcel
I/O aware multi-threading
<defmod name=“CORBA-omniORB-4.0” driver=“binary”>
<attr label=“NameService”> corbaloc::paraski.irisa.fr:8088/NameService </attr>
<requires>SysW</requires>
<requires>VIO</requires>
<requires>LibStdC++</requires>
<unit>CORBA-omniORB-4.0.so</unit>
<file>lib/libomniorb-4.A.so</file>
<file>lib/libomnithread4.A.so</file>
</defm=od>
EuroPVM/MPI 03 - Sept., 2003
Programming High Performance Applications using Components
HLA
PadicoTM
Core
PadicoTM
Services
SOAP
or
CORBA
Create
Load
Start
Stop
Unload
Destroy
Multithreading
XML
Module
Descriptor
23
Performances

Experimental protocols



Runtimes and middleware:
MPICH, CORBA OmniORB3, Kaffe
JVM
Hardware: Dual-Pentium III 1 Ghz
with Myrinet 2000, Dolphin SCI
Performance results

250
CORBA/Myrinet-2000
MPI/Myrinet-2000
200
Java/Myrinet-2000
CORBA/SCI
MPI/SCI
TCP/Ethernet-100
150
Myrinet-2000




Bandwidth (MB/s)
MPI: 240 MB/s, 11 ms
CORBA: 240 MB/s, 20 ms 18 ms
96 % of the maximum achievable
bandwidth of Madeleine
100
50
SCI


MPI: 75 MB/s, 23 ms
CORBA: 89 MB/s, 55 ms
0
1
10
100
1000
10000 100000 1000000 1E+07
Message size (bytes)
EuroPVM/MPI 03 - Sept., 2003
Programming High Performance Applications using Components
24

Composition scalability



GridCCM code hand written
1D block distibution
Experimental Protocol

Platform



GridCCM implementation



16 PIII 1 Ghz, Linux 2.2
Fast-Ethernet network + Myrinet-2000 network
JAVA: OpenCCM
C++: MicoCCM
C++/Myri based on
MicoCCM/PadicoTM
Comp. A-0
Comp. A-1
A-0
SPMD
Comp. A-2
A-0
Proc.
SPMD
Comp. A-3
A-0
SPMD
Comp. A-4
A-0
Proc.
SPMD
Proc.
SPMD
Proc.
Proc.
EuroPVM/MPI 03 - Sept., 2003
Comp. A-0
Comp. A-1
A-0
SPMD
Comp. A-2
A-0
Proc.
SPMD
Comp. A-3
A-0
SPMD
Comp. A-4
A-0
Proc.
SPMD
Proc.
SPMD
Proc.
Proc.
Aggregated Bandwidth in MB/s
Preliminary GridCCM performance
300
250
200
150
100
50
0
1->1
2->2
4->4
8->8
Component configuration
Java
C++/Eth
C++/Myri
Programming High Performance Applications using Components
25
Conclusion & Future works

Conclusions

Code coupling tools can benefit from component models


An existing component model such as CCM can be extended to comply with
HPC requirements without modifying the standard


Provide a standard to let HPC codes to be reused and shared in various contexts
We do not need to reinvent the wheel…
CORBA is a mature technology with several open source implementations

Should be careful when saying that CORBA is not suitable for HPC




It is a matter of implementation ! (zero-copy ORB exists such as OmniORB)
It is a matter of a suitable runtime to support HPC networks (PadicoTM)
Yes, CORBA is a bit complex but do not expect to have a simple solution to program
distributed systems…
Future works

Deployment of components using Grid middleware & dynamic composition
EuroPVM/MPI 03 - Sept., 2003
Programming High Performance Applications using Components
26
Some references

The Padico project


One of the best tutorial on CCM


http://www.irisa.fr/paris/Padico
http://www.omg.org/cgi-bin/doc?ccm/2002-04-01
A web site with a lot of useful information about CCM

EuroPVM/MPI 03 - Sept., 2003
http://ditec.um.es/~dsevilla/ccm/
Programming High Performance Applications using Components
27