parallel_armadillo

Benchmark code and results for parallelizing the diagonalization of matrices using armadillo.

Dependencies

Goal

The goal of this project is to find out some timings of full diagonalizations of large, symmetric matrices using the armadillo library, on my machine. Specifications:

• MacBook Pro, mid 2015
• Processor: 2,5 GHz Intel Core i7 (4 cores, 256KB L2, 6 MB L3)
• 16 GB 1600 MHz DDR3
• Intel Iris Pro 1536 MB VRAM

C++ Compiler:

• (MacPorts gcc49 4.9.4_2) 4.9.4

Timings

Preliminary testings

With standard armadillo, we get the following timings (in seconds). Armadillo automatically enables OpenMP:

50 : 0.00419
100 : 0.001029
250 : 0.004986
500 : 0.021301
1000 : 0.099476
2500 : 1.44927
5000 : 13.0289
10000 : 164.181

-01 compiler flag

Using the -01 compiler flag, we get the following timings:

50 : 0.000217
100 : 0.000885
250 : 0.005373
500 : 0.022267
1000 : 0.106715
2500 : 1.59232
5000 : 13.3628
10000 : 170.674

-O2 compiler flag

Using the -O2 compiler flag, we get the following results:

50 : 0.000223
100 : 0.001011
250 : 0.005317
500 : 0.021279
1000 : 0.106962
2500 : 1.58473
5000 : 12.8927
10000 : 164.331

-O3 compiler flag

Using the -O3 compiler flag, we get the following results:

50 : 0.000211
100 : 0.000993
250 : 0.004822
500 : 0.021878
1000 : 0.098428
2500 : 1.65663
5000 : 13.3471
10000 : 162.549

Accelerate and -O2

Using the Accelerate (macOS) framework, we get the following timings:

50 : 0.00021
100 : 0.000957
250 : 0.004876
500 : 0.021395
1000 : 0.096768
2500 : 1.53004
5000 : 12.3652
10000 : 158.362

Sparse matrices & Accelerate and -O2

1000 : 0.177537
2500 : 1.2949
5000 : 8.53172
10000 : 85.7126