Skip to content

gcoe-dresden/cuda-gpu-tlb

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

16 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

CUDA GPU TLB Benchmarks

Micro-Benchmarks for Discovering TLB Cache Level Hierarchies.

Requirements

  • cmake 2.8+
  • C++ compiler (tested with gcc5.3.0)
  • CUDA (7.5, 8.0 or newer)

Build

Go to the cuda-gpu-tlb directory (created by git clone ...):

mkdir release && cd release
cmake -DM_CUDA_ARCH=30 .. # or cmake .. && ccmake . # to configure device code
make -j 2

tlb-bench Usage

$ ./tlb-bench 
usage: ./tlb-bench data_from_MB data_to_MB stride_from_KB stride_to_KB Device_No=0 min_instead_avg=0

Test for the K80 - generates a CSV file

./tlb-bench 1 5 64 256
./tlb-bench 48 300 1024 4096
./tlb-bench 1500 5000 1024 4096

Plotting needs R - generates a pdf file

cd supplemental/
chmod u+x tlb-benchmark-plot.r
./tlb-benchmark-plot.r TLB-Test-1-5-64-256.csv
./tlb-benchmark-plot.r TLB-Test-48-300-1024-4096.csv
./tlb-benchmark-plot.r TLB-Test-1500-5000-1024-4096.csv

or just use the provided Makefile.

Known issues

Kepler and Pascal GPUs seem to work fine but we had some issues getting good results on Maxwell GPUs. This is future work.

TLB Sharing

tlb-sharing usage

./tlb-sharing 
usage: ./tlb-sharing stride_KB iterations device_No=0 min_instead_avg=0
  • device_No is your device ID
  • min_instead_avg uses minimum of the results of the benchmark iterations

try:

  • stride_KB = page_size and
  • iteration = #entries
  • choose the parameters that (otherwise you will not see the wanted effects):
stride_kb * iterations < TLB but 2 * stride_kb * iterations > TLB

output should be something like this:

./tlb-sharing 2048 65
#Tesla K80: cuda 3.7
#----------- absolute values ---------------
# 0 1 2 3 4 5 6 7 8 9 10 11 12 
0 342 286 285 286 285 343 285 286 285 287 285 286 285 
1 293 346 287 287 287 286 346 287 288 287 287 286 287 
2 285 281 336 281 280 281 280 337 280 281 336 281 280 
3 286 280 281 337 281 280 281 280 338 280 282 337 281 
4 286 282 281 282 336 282 281 282 281 337 281 282 336 
5 348 290 291 291 291 347 291 290 291 290 291 290 291 
6 297 352 292 293 291 293 351 293 291 293 291 293 291 
7 290 284 341 284 285 284 285 340 285 284 341 284 285 
8 290 285 284 342 284 285 284 285 341 285 284 342 284 
9 291 285 286 285 341 285 286 285 286 340 286 285 341 
10 294 290 345 290 289 290 289 346 289 290 345 290 289 
11 296 289 290 346 290 289 290 289 347 289 290 346 290 
12 295 291 290 291 345 291 290 291 290 346 290 291 345 
#----------- which SMs interfere ---------------
# 0 1 2 3 4 5 6 7 8 9 10 11 12 
0 .X .. .. .. .. .X .. .. .. .. .. .. .. 
1 .. .X .. .. .. .. .X .. .. .. .. .. .. 
2 .. .. .X .. .. .. .. .X .. .. .X .. .. 
3 .. .. .. .X .. .. .. .. .X .. .. .X .. 
4 .. .. .. .. .X .. .. .. .. .X .. .. .X 
5 .X .. .. .. .. .X .. .. .. .. .. .. .. 
6 .. .X .. .. .. .. .X .. .. .. .. .. .. 
7 .. .. .X .. .. .. .. .X .. .. .X .. .. 
8 .. .. .. .X .. .. .. .. .X .. .. .X .. 
9 .. .. .. .. .X .. .. .. .. .X .. .. .X 
10 .. .. .X .. .. .. .. .X .. .. .X .. .. 
11 .. .. .. .X .. .. .. .. .X .. .. .X .. 
12 .. .. .. .. .X .. .. .. .. .X .. .. .X 

Publication

The benchmark approach and results for the Nvidia K80 and P100 are described in:

Big Data causing Big (TLB) Problems: Taming Random Memory Accesses on the GPU

Tomas Karnagel, Tal Ben-Nun, Matthias Werner, Dirk Habich, and Wolfgang Lehner -- Thirteenth International Workshop on Data Management on New Hardware (DaMoN 2017)

http://dx.doi.org/10.1145/3076113.3076115