Skip to content

karnajitsen/NVLink-benchmark

Repository files navigation

Configuration:

/config/parameters.config file contains all the parameters of measurement to be configured. In our parameter file, we have used following letters to denote unit of data size, B = Byte K = Kilo Bytes M = Mega Bytes G = Giga Bytes

An example of parameters and there description are as follows,

TPDATATYPE = double : data type for througput experiement LATDATATYPE = int : data type for latency experiement SIZE = 32M-8G,2 : Size of the array, data size. format - 32 = Starting minimum data size, 8 = Last maximum data size, "-" = Minimum and maximum data size separator, "," = data size and incrementing factor separator, "2" = incrementing factor multiple of 2, ACCESSPATTERN = LINEAR : Access pattern, not used so far THROUGHPUT = N : "Y" = throughput experiment will takes place, if "N" no throughput experiements CUDAMEMCOPY = N : "Y" = cudamemcpy experiment will takes place, if "N" no cudamemcpy experiements KER2MEMLOC = C0-C0;C0-G0;G0-G1; C0-G0-G1 : "Kernel location"-Memory location 1, Memory Location 2, ... ", Mapping Separator is ";" STREAM = READ,WRITE,COPY,ADD,TRIAD,QUADAD,PENTAD,HEXAD : STREAM type: One can remove any of these stream, but the order should be maintaned as of now. MEMORYMODE = NONUM : Memory models , NONUM = Zero-Copy , UM = Unified Memory REPETATIONS = 1 : # of Repeatations for the experiments to avoid measurement noise LATENCY = Y : Latency experiements will take place if "Y" else no latency experiements. Both THROUGHPUT and LATENCY flag can not be "Y" at the same time. ITERATIONS = 1 : # of iterations for pointer chasing, kept 1 by default for our experiments STRIDE = 4B-512K,2 : Starting minimum stride size - Last Maximum stride size, Incrementing value, Incrementing operator ( * - multiplications, + - Addition ) CHASE_ELEMENTS = 128-16384,2,* : Starting minimum # of elements chased - Last Maximum # of elements chased, Incrementing value, Incrementing operator ( * - multiplications, + - Addition ) START_THREAD = 1024 : Starting gpu block size END_THREAD = 1024 : Maximum GPU block size OMP_THREADS = 1-1,4 : Starting minimum Openmp thread - Last Maximum Openmp threads, Incrementing value to be multiplied MEM_ADVISE = N : cuMemAdvise() experiment is performed if "Y" with MEMORYMODE = UM , if "N" , it will not take place. THREAD_EXPERIMENT = N : Reduction of Redudant page fault experiments will take place if "Y" with MEMORYMODE = UM else no.

Execution:

Execution can be done executing simple make command after correct configuration. Different make commands are,

  1. make nvl0 : Complete compilation and execution of the experiments for all the range of data size given in configuration file for NVLINK system.
  2. make pcie0 : Complete compilation and execution of the experiments for all the range of data size given in configuration file for PCIe system.
  3. make cmp0: Only compilation for NVLINK System.
  4. make cmp1: Only compilation for PCIe system.

For individual execution, one needs to run following command,

./nvprog :

A : Data size for throughput / # of chasing elements for latency B : Stride length in elements C : # of Open mp threads D : 1 for throughput , 0 for latency

Output:

After execution, output will be stored inside /data/ folder as csv and text file with following naming conventions.

  1. result-pchase-lat-nonum.txt or result-pchase-lat-nonum.csv - for zero copy latency experiment.
  2. result-pchase-lat-um.txt or result-pchase-lat-um.csv - for Unified memory atency experiments.
  3. result-stream-bw-nonmum.txt or result-stream-bw-nonmum.csv - zero-copy memory throughput output.
  4. result-stream-bw-um.txt or result-stream-bw-um.csv - Unified memory throughput output.
  5. result-stream-bw-memcpy.txt or result-stream-bw-memcpy.csv - cuda mem copy output.

For complete execution, data file from previous executions are always backed up.

About

NVLink microbenchmark with IBM Power8 and NVIDIA P100 GPU - Master Thesis

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published