Middleware for Grid computing: experiences using the VTHD Network (2.5 Gbit/s)

Transcription

1 Middleware for Grid computing: experiences using the VTHD Network (2.5 Gbit/s) Frédéric DESPREZ ENS-Lyon/INRIA & Thierry PRIOL IRISA/INRIA Contents: - The RNRT VTHD project - How to exploit a high bandwidth network transparently? - High-performance distributed objects platform (Paco/PADICO) - Global address space for the Grid (MOME) - Computational servers on a High Speed WAN (DIET) October 26, 2001 Grid computing experiences with VTHD 1

2 VTHD project, Vraiment Très Haut Débit Coordinator: Christian Guillemot (FTR&D) October 26, 2001 Grid computing experiences with VTHD 2

3 Project organisation SP1: Platform deployment (FTR&D, INRIA, ENST,ENST-Br, INT, EURECOM) IP engineering and WDM, Back-office IP, «best-effort» IPv4 service SP2: Flow and congestion control (INRIA, FTR&D) TCP, Web traffic genetation and monitoring (WAGON/ INRIA software). SP3: Multi-service high-bandwidth platform (ENST, ENST-Br, INT, FTR&D) QoS service class, RPV service, Tera-routers. SP4: Training and medical applications (ENST, INRIA, ENST-Br, INT) Contents server, remote training (ENST-Br,INT,ENST). Medical «white board», Data transmission associated with the use of medical robots (INRIA, HEGP). SP5: Distributed computing and visualisation resources (INRIA) Grid computing, distributed numerical simulation, remote visualisation. SP6: Distributed cache systems (EURECOM, ENST, INRIA) Data localisation, benchmarking WAGON. October 26, 2001 Grid computing experiences with VTHD 3

4 VTHD network Sycamore SN16000 Cisco GSR Junipe r M40 Avici TSR Access Router IPv6 router (Cisco 7200) Lannion FTR&D FTR&D INRIA ENST-Br FTR&D Caen Rennes Rouen Paris Nancy INRIA WDM network with «packet over SDH» channels 2,5 Gbit/s. IPv4 network supported by 8 giga-routers Renater IPv6 Nantes FTR&D E-BGP Xantium Atrium FT HEGP INRIA Grenoble Sophia Antipolis INRIA FTR&D ENST INT INRIA FTR&D EURECOM 17 access routers to aggregate local traffic network over Giga-Ethernet fiber IPv6 network interconnected by IPv6/GEth/MPLS or IPv6/IPv4 tunnels October 26, 2001 Grid computing experiences with VTHD 4

5 VTHD computational grid Inria Rhône-Alpes HP Cluster nodes Inria Rennes 34 nodes cluster SCI / Myrinet / Gigabit Ethernet October 26, 2001 Inria Sophia-Antipolis 16 nodes cluster Grid computing experiences with VTHD 5

6 Context of the experiments At the beginning of the project «2.5 gigabits is probably too much for PCs that are connected to the VTHD network with 100 Mbits/s Ethernet interfaces» Goal of the research works Grid = cluster of cluster Design middleware that allow transparent exploitation of high-bandwidth network For scientific applications (parallel codes) Principle Exploit data distribution within a parallel code to exploit high-bandwidth networks October 26, 2001 Grid computing experiences with VTHD 6

7 Distributed objects oriented platforms Distributed objects / Components To enforce users to structure their applications Parallel codes encapsulation Coupling Through a scalable middleware Structural mechanics Themal Optics Dynamic Sequential componentl Parallel Component SAN LAN middleware Homogeneous cluster SAN WAN Sequential component Parallel component Homogeneous cluser Visualisation Supercomputer October 26, 2001 Grid computing experiences with VTHD 7

8 Mapping of components onto resources WAN SAN SAN WAN LAN To deploy the same code but using various networking resources SAN Configuration 1 Scalable Middleware SAN Paco / Paco++ High-Performance Middleware Padico Configuration 2 The middleware should adapt itself to the available networking resources October 26, 2001 Grid computing experiences with VTHD 8

9 CORBA : a middleware for scientific computing CORBA: Common Object Request Broker Architecture 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) Problems to face Performance issues (not true) Poor integration of high performance computing environments interface MatrixOperations { const long SIZE = 100; typedef double Vector[ SIZE ]; typedef double Matrix[ SIZE ][ SIZE ]; void multiply( in Matrix A, in Vector B, out Vector C ); }; Client Obj. Inv. IDL Stub IDL Compiler Obj. Impl. IDL Skel. OA Object Request Broker (ORB) Server October 26, 2001 Grid computing experiences with VTHD 9

10 Goal Parallel CORBA object concept Encapsulation of parallel codes into CORBA objects Scalable connection between parallel CORBA objects Supercomputer interface[*:2*n] MatrixOperations { const long SIZE = 100; typedef double Vector[SIZE]; typedef double Matrix[SIZE][SIZE]; void multiply(in dist[block][*] Matrix A, in Vector B, Machine A Client Extended -IDL specification for MPI-based parallel codes Extended-IDL compiler Object impl. SPMD code Skel. Parallel CORBA object MPI Communication layer Object impl. SPMD code Skel. Object impl. SPMD code Skel. Object impl. SPMD code Skel. out dist[block] Vector C ); void skal(in dist[block] Vector C, Stub PBOA PBOA PBOA PBOA }; out csum double skal); CORBA ORB Submitted to the OMG as a response to a RFI October 26, 2001 Grid computing experiences with VTHD 10

11 VTHD Experiment q Protocols Ethernet 100 Mb/s u Two MPI codes u Encapsulation using PaCO++ q VTHD Performance 1 Gb/s Switch u 11 nodes to 11 nodes u 826 Mb/s (103 MB/s) u P2P: 75 Mb/s (9.4 MB/s) u Near the switch limit VTHD 2.5 Gb/s q Cluster Performance u 8 nodes to 8 nodes (Myrinet 2000) u 12 Gbit/s (1.5 GB/s) u P2P: 1.5 Gbit/s (187 MB/s) Switch 1 Gb/s Ethernet 100 Mb/s October 26, 2001 Grid computing experiences with VTHD 11

12 Cluster performance PADICO: an Open Integration platform for communicating middleware and runtimes October 26, 2001 Grid computing experiences with VTHD 12

13 Global address space for the Grid (MOME) Goal To transfer data between DSM-based parallel codes Distributed Shared Memory coupling Parallel code Code B Distributed Shared Memory Code A Coupling Interface Cluster A Cluster B October 26, 2001 Grid computing experiences with VTHD 13

14 VTHD Experiment q Protocol u Three codes coupled with MOME (Simulation + visualisation) q VTHD Performance u 11 nodes to 11 nodes u 850 Mb/s (106 MB/s) Ethernet 100 Mb/s 1 Gb/s Switch Ethernet 100 Mb/s VTHD 2.5 Gb/s Switch 1 Gb/s Switch 1 Gb/s Ethernet 100 Mb/s October 26, 2001 Grid computing experiences with VTHD 14

15 The JACO3 Environment A set of CORBA services to support coupled simulation applications Standalone vis. application Web-based vis. applications User management, concurrent application execution, sharing of results Data path = FTP Control path = CORBA IIOP JACO3 Software bus Visualisation tools Application Control Panel User & Session Service Activation Service Application deployment and destruction, hiding of component distributio JACO3 Software Bus Code coupling tool Application scheduler Controls User s application Application components Data Repository Service Computing Resource Allocation Service Information Repository Service Data flow between components and input for visualisation Batch resource and coallocation management Application, component, resource and user information VisualORB LDAP DB October 26, 2001 Grid computing experiences with VTHD 15