IBM MASS, Turbo MPI, and MIO libraries

1
IBM MASS, Turbo MPI, and MIO libraries

• Bob Sinkovits
• Scientific Computing Department
• SDSC
• All hands Meeting
• February 24-28, 2001

2
What is MASS?

• Mathematical Acceleration SubSystem (MASS) is a
  library containing high performance versions of a
  subset of the Fortran intrinsic functions
• Can be used with Fortran or C
• Applications that make heavy use of intrinsic
  functions, can get a big speedup using MASS

3
How to use MASS

• No source code modifications required. Calls to
  intrinsic functions automatically replaced with
  calls to the MASS library.
• No need to recompile, simply link
  with -L/usr/local/apps/mass -lmass

4
MASS Accuracy considerations

• The MASS library achieves the improved
  performance at the expense of a small amount of
  accuracy.
• The difference between MASS and the standard math
  library (libm.a) is restricted to the last bit of
  the mantissa and should be insignificant for most
  codes.
• If using MASS makes a noticeable difference in
  results, there is probably a problem with your
  algorithm

5
Vector MASS

• The MASS library also contains a set of vector
  intrinsic functions that can yield even better
  performance.
• In a vector version of the intrinsic function,
  operations are pipelined. Can overlap the
  evaluation of the function on successive elements
  of the vector.

Z(3)exp(x(3))
Z(4)exp(x(4))
Z(2)exp(x(2))
Z(1)exp(x(1))
Z(1)exp(x(1))
Z(2)exp(x(2))
Z(3)exp(x(3))
Z(4)exp(x(4))
6
Vector MASS

• Vector mass routines can be used whenever an
  intrinsic function needs to be applied to a
  vector of variables. The most obvious case is a
  simple loop containing a call to the intrinsic
  function.

!Perfect candidate for replacement with vector
call do i 1,n z(i) exp(x(i)) enddo
7
Using Vector MASSCoding by Hand

• Loops over calls to intrinsic routines are
  replaced with calls to vector mass routines at
  the source code level do i1,n z(i)
  exp(x(i)) ? call vexp(z,x,n) enddo
• Need to explicitly link vector mass
  library-L/usr/local/apps/mass -lmassvp3

8
Using Vector MASSLet the compiler do the
substitution

• Loops over calls to intrinsic routines are
  replaced with calls to vector mass routines at
  the assembly code level. To verify that
  substitution took place, look at the assembly
  code (generated with -S compiler option) do
  i1,n z(i) exp(x(i)) ? bl .__vexpPR
  enddo
• Need to compile with global optimization level of
  -O5

9
MASS and Vector MASS performance
Time for 106 function evaluations
10
Useful MASS URLs

• Vector intrinsic functions II IBM SP(This
  content is slightly out of date for XLFortran
  v7.0)www.npaci.edu/online/v3.13/SCAN1.html
• Vector intrinsic functions I CRAY
  T3Ewww.npaci.edu/online/v3.9/SCAN1.html
• IBM optimization libraries (including
  MASS)www.research.ibm.com/actc/Opt_Lib/Topic_OptL
  ibraries.html

11
Turbo MPI

• Turbo MPI library contains a subset of the MPI
  collective operations that have been optimized
  for the clustered SMP architecture
• MPI_BARRIER
• MPI_ALLREDUCE
• MPI_REDUCE
• MPI_BCAST
• MPI_ALLTOALL

12
Turbo MPI BCAST example

• Step 1 N0 ? N1
• Step 2 N0 ? N2 N1 ? N3
• Step 3 Shared memory operations within each node

13
Turbo MPI

• Turbo MPI is not an IBM supported product, Use
  at your own risk!
• Initial evaluations of Turbo MPI indicate that it
  is slower than the standard IBM MPI library.
  Benchmark your application carefully!
• To use Turbo MPI, link with-L/usr/local/apps/turb
  oMPI -lTurboMPI
• Documentation available on Blue Horizon
  at/usr/local/apps/turboMPI/turbo.ps

14
MIO (Modular I/O)

• MIO is an unsupported I/O optimization library
• We have very recently received MIO and will be
  installing shortly on Blue Horizon
• MIO library available on Blue Horizon
  at/usr/local/apps/MIO/libmio.a
• Documentation on MIO available atwww.research.ibm
  .com/actc/Opt_Lib/mio/mio_doc.htm