Title: Scalable Linux MultiComputer
1Scalable Linux MultiComputer
- IEEE RADAR CONFERENCE
- April 2002
2Overview
- Introduction
- Hardware Architecture
- Evolution
- Interprocessor Communication
- Strategic Directions In HPEC
- Embedded Linux
- Conclusions
3Hardware Architecture
- First Level - I/O Processor
- LANai Network Processor off loads the PPC
- Myrinet Scalable Cluster Interconnect (parallel
and serial) - Concurrent I/O and processing
- Second Level - Compute Processor
- PowerPC AltiVec architecture
- Low Power Consumption
- High Performance per Watt
- AltiVec SIMD vector processing canyield from 2
to 10X performance increase over FPU systems. -
4System Feature Evolution
- Development and Deployment Platforms now use
same technology - Based on Open Standards and Open Source Software
- Support for PowerPC processors running VxWorks or
Linux - Improved Systems Scalability ( 6.4 GFLOPS to 3.2
TFLOPS) - Processing Node Migration from PowerPC to
AltiVec architecture - Enhancements to Fast Boot Capabilities
- Kernel Boot, NFS mount
- Card is fully operational with OS 10 seconds
after insertion - Power On Self Test
- Fault Tolerant Features Added
- Live Insertion - Hot Swap (VME handles have power
switch) - Internal current and thermal sensors for
monitoring automatic shut down - Fault Tolerant API Library
- New Technology Insertion to Scalable Cluster
Interconnect - Dynamic or Static Mapping of the Network
Interconnect (Myrinet) - 8 Port Switch upgraded to 16 port Switch
- Serial 2 Gb technology over Fiber for chassis to
chassis interconnect
5Performance Evolution PowerPC Node
6Performance Evolution Myrinet Interconnect
7Backplane Rugged Fabric 24x 16 Port Switches
16x 6U Quad DSP Boards 32x Fiber I/O ports
8System Rugged Fabric 16 Chassis - 1024x
DSP(AltiVec) Nodes
9Strategic Directions inHigh Performance Embedded
Computing
Adherence to Standards
Open Source Software
Proprietary
APPLICATIONS
Reuse is High
Reuse is Low
MIDDLEWARE
Compatibility is Low
Compatibility is High
DEVELOPMENT TOOLS
Availability is Low
Availability is High
OPERATING SYSTEM
Home Grown OS Development/Support Maintenance
Costs High
VxWorks and Linux Development/Support
Maintenance Costs Low
PROCESSOR/HARDWARE TECHNOLOGIES
10Linux for the Embedded Processor
- Small kernel and boot file system in EEPROM, with
space for application. - Kernel can be compiled to include only necessary
drivers for embedded hardware and application. - Dynamically loaded device drivers is automatic.
- ROM compressed kernel and root tools
- RAM temporary, process directories
- NFS common set of config, user home reside
on host disk system - Zero size swap space to disable virtual memory
11Linux for the Embedded Processor
- Fast restart time in mission-critical
applications - 32 MB FLASH EEPROM
- Power On Self Test prior to booting OS
- Boot Compressed Kernel from FLASH
- Read Static Network Routes from FLASH
- Load User Application from FLASH
- System Boots OS in less than 10 Seconds
- FLASH management S/W routines available
- Familiar embedded environment, node looks just
like a desktop processor. - Uses shells, scripts, standard /etc config
files, and NFS mounted disks. - Common and open tools gcc, gdb, sh, cvs, X,
MPI, MPE.
12Data Passing Methods
- Myrinet-2000
- 250 MB/s peak
- Portable open GM Interface
- Lowest Level interface API
- TCP/IP
- Sockets semaphores
- Easy interface to other computers via PCI card
- MPI
- Portable open MPICH
- High level interface API
- Large developer base for latest advanced features
- Supports TCP/IP, GM, BDMP devices
- CORBA
- Object Oriented API
- TCP/IP, optimized with GM
13Conclusion
- CSPI offers MultiComputers ready for Mission
Critical COTS Applications - The Software development environment is in place
- The performance (Linux VxWorks) is comparable
and scalable - Systems are based on industry standards (MPI,
VSIPL for AltiVec, VxWorks Linux for MPC7410) - Systems optimize processing density, scalability,
and bi-section bandwidth - Systems meet the needs of a wide range of DOD
signal processing applications (radar, sonar,
surveillance and C3I)Further Information
available at - http//www.cspi.com/multicomputer/
- http//www.myri.com