Effective Bandwidth (b_eff) Benchmark - Implementation Details

Implementation Details of b_eff (version 3.3)

• The measurements are done in nested loops:
  loop over all repetions
  loop over all message lengths
  loop over all patterns
  loop over all methods
  a barrier
  one message exchange (except for the 3 largest message lengths)
  time_begin = MPI_Wtime()
  loop over looplength
  message exchange
  time_end = MPI_Wtime()

• Two buffers are used, one for sending and the other for receiving the messages, each of size 6*(L_max+16).

• The messages are allocated cyclically in the buffers to inhibit reading from the cache.

• Before sending a message, a value is assigned to the first and last byte. After receiving a message, always these two bytes are checked. If an error is detected the benchmark is aborted immediately and a detailed error report is written to standard output.

• If the estimation for the execution time for the two longest message lengths will be more than 60 seconds, then only the fastest method (based on the third longest message length) is used and the number of repetitions is reduced.

• On a system with a latency less than 50 micro sec and nearly same bandwidth for all pattern and methods, the total execution time is about
  1..4 * max( 60 sec, 3*MEMORY_PER_PROCESSOR/asymptotic bandwidth )

UP     Effective Bandwidth Benchmark     Pallas Effective Bandwidth Benchmark     MPI at HLRS     HLRS Navigation     HLRS    

Rolf Rabenseifner