Parallel/Concurrent Programming on the SGI Altix

1
Parallel/Concurrent Programming on the SGI Altix

• Conley Read
• January 25, 2007
• UC Riverside, Department of Computer Science

2
Overview

• The SGI Altix
• Itanium 2 Features
• Tools
• Schedutils
• MPI Programming
• Star-P
• References
• Try It

3
About the SGIs

• sgi-2.cs.ucr.edu
• 32 processor
• Intel Itanium
• 1.5 GHz
• 128 GB (70 GB) RAM
• sgi-dev.cs.ucr.edu
• 4 processor (2 sockets X 2 cores)
• Intel Xeon
• 3 GHz
• 8 GB RAM
• 2 Virtex 4 LX200 FPGAs
• 200,000 Logic Cells
• 80 MB SRAM

4
SGI Altix 3000

• Runs one copy of Fedora or OpenSUSE Linux
• Ported to SGI Hardware (Lessons from IRIX)
• Linux natively supports management of 32
  processors
• 1 copy of Linux vs. 1 copy of Linux per node
• Node sizes 32,64,128?
• Programming for the 32 processor SGI is similar
  to programming for a 4 way SMP

5
Global Shared Memory

• Single Memory space
• All system resources
• Processors
• I/O
• SGI Altix 3000 uses NUMAflex (interconnect) to
  connect Itaniums with NUMAlink (ASIC)
• Creates SMP-like system
• Regardless of the node location of thread
  execution
• Interprocess communication time is reduced

6
The Itanium 2

• First Supercomputer using Itanium (1)
• IBM Titan NOW Cluster at NCSA
• Unique Features
• 16 Byte Atomic Operations
• IA-64
• Fixed Length Instructions
• Relies on Compile-time Optimizations
• Branch Speculative Execution
• Speculatively fetches data not in Cache
• Virtualization (Intel Vanderpool Technology or
  VT)
• Explicitly Parallel Instruction Computing (EPIC)

7
Tools

• GNU C Compiler Installed
• Intel C Compiler (icc) Waiting for License
• Intel Debugger (idb)
• GNU gdb Debugger Installed
• Some problems with code generated by icc
• Etnus TotalView Better Debugger
• Intel Vtune Available for purchase
• sTrace System Call Tracer
• SGI Star-P Currently being installed

8
ICC

• Intel C Compiler (icc)
• OpenMP Simpler conversion from serial to
  parallel execution
• Auto-parallelization Safe-loop parallelization
• Fortran 95
• Optimizations for IA-64 Architecture
• Itanium 2 uses IA-64

9
Schedutils

• Change Process Scheduling Attributes
• chrt options pid command
• -p Edit existing PID
• --rr Round Robin
• --fifo SCHED_FIFO
• --other SCHED_OTHER

10
Schedutils

• Change Process CPU Affinity
• taskset options mask pid command
• -p Edit existing PID
• Mask
• 0x00000001 Processor 0 0001
• 0x00000002 Processor 2 0010
• 0x00000003 Processor 0 and 1 0011
• 0xFFFFFFFFF All Processors 1111

11
Parallel Programming Building Blocks

• SGI Altix supports common MPI libraries
• Pthreads not exactly threads
• SGI Message Passing Interface
• Intel MPI Library
• Voltaire MPI Installed
• Scali MPI

12
MPI Programming

• Create a new thread
• fork() vs. pthreads_create()
• Hardware vs. POSIX
• Bind a thread to a processor
• pthread_processor_bind_np()
• If you do not bind the pthread or modify the
  forked thread, the new thread will inherit the
  scheduling properties of the creating thread.

13
Star-P

• The power of Supercomputers
• MATLAB to SGI Altix over Network
• Offloads computation tasks
• Runtime call intercept
• Perform day long computations in minutes
• Without the Hassle
• No reprogramming in C MPI

14
How to Login

• SGI Altix at UCR
• sgi-2.cs.ucr.edu
• Use secure SSH protocol to connect
• CS Username
• CS Password
• SGI RASC Altix at UCR
• sgi-dev.cs.ucr.edu
• Use secure SSH protocol to connect
• CS Username
• CS Password

15
Try it

• ssh sgi-2.cs.ucr.edu
• cat /proc/cpuinfo
• cat /proc/meminfo
• free
• vmstat
• uname -a

16
Questions?
17
References

• http//www.intel.com/products/processor/itanium2/
• http//en.wikipedia.org/wiki/Itanium2
• http//www.sgi.com/developers/technology/gsm.html
• http//rlove.org/schedutils/
• http//www.llnl.gov/computing/tutorials/pthreads/

18
Thank you!