Performance Examples

[Laurae2]

OLD 2^31-1 BN

Server:

• Dual Xeon Gold 6130 (2x 16 cores / 32 threads, 2.8 GHz all turbo, 3.7 GHz singlethread)
• 384GB RAM (12 x 32GB RAM, 2666 MHz)
• R 3.5.0 compiled using gcc-5.4, -O3 -mtune=native flags
• Linux Subsystem for Windows, Ubuntu 16.04
• BIOS additional settings: UMA mode (not NUMA)

CPU frequency:

Frequency	CPU loaded
3.7 GHz	1, 2
3.5 GHz	3, 4
3.4 GHz	5, 6, 7, 8
3.1 GHz	9, 10, 11, 12
2.8 GHz	13, 14, 15, 16
2.8 GHz	17-32

On a 16GB vector (2^31 - 1 elements), parallel mean:

What	Threads	Elapsed Time	CPU Time	Throughput	Information
R	1	6.137s	6.141s	0.3 bn/s	Handles NA. Handles more than 2^31 - 1 elements.
C++	1	3.147s	3.147s	0.7 bn/s	No checks on data.
C++	2	1.613s	3.206s	1.3 bn/s	No checks on data.
C++	4	0.832s	3.304s	2.6 bn/s	No checks on data.
C++	8	0.432s	3.432s	5.0 bn/s	No checks on data.
C++	16	0.229s	3.664s	9.4 bn/s	No checks on data.
C++	32	0.172s	5.241s	12.4 bn/s	No checks on data.
C++	61	0.152s	9.102s	14.0 bn/s	No checks on data. Optimal run.
C++	64	0.165s	9.791s	13.0 bn/s	No checks on data.

[Laurae2]

NEW 10 BILLION

Server:

• Dual Xeon Gold 6130 (2x 16 cores / 32 threads, 2.8 GHz all turbo, 3.7 GHz singlethread)
• 384GB RAM (12 x 32GB RAM, 2666 MHz)
• R 3.5.0 compiled using gcc-5.4, -O3 -mtune=native flags
• Linux Subsystem for Windows, Ubuntu 16.04
• BIOS additional settings: UMA mode (not NUMA)

CPU frequency:

Frequency	CPU loaded
3.7 GHz	1, 2
3.5 GHz	3, 4
3.4 GHz	5, 6, 7, 8
3.1 GHz	9, 10, 11, 12
2.8 GHz	13, 14, 15, 16
2.8 GHz	17-32

On a 80GiB (74.5GB) vector (10,000,000,000 elements), parallel mean:

> set.seed(1)
> my_vec <- runif(n = 1e10, min = -0.5, max = 1)
> object.size(my_vec)
80000000048 bytes

Repeats per number of threads from 1 to 64:

# Approx 4521.241 seconds total
> ceiling(20 * log(seq_len(64) + 1))
 [1] 14 22 28 33 36 39 42 44 
 [9] 47 48 50 52 53 55 56 57
[17] 58 59 60 61 62 63 64 65
[25] 66 66 67 68 69 69 70 70
[33] 71 72 72 73 73 74 74 75
[41] 75 76 76 77 77 78 78 78
[49] 79 79 80 80 80 81 81 81
[57] 82 82 82 83 83 83 84 84

Results of the parallel mean:

What	Threads	Elapsed Time	CPU Time	Throughput	Speedup vs R	Information
R	1	33.235s	33.235s	0.3 bn/s	1x	Handles NA.
C++	1	16.139s	16.141s	0.6 bn/s	2.06x	No checks on data.
C++	2	7.985s	15.527s	1.3 bn/s	4.16x	No checks on data.
C++	4	4.024s	15.674s	2.5 bn/s	8.26x	No checks on data.
C++	8	2.101s	16.188s	4.8 bn/s	15.82x	No checks on data.
C++	16	1.097s	16.782s	9.1 bn/s	30.30x	No checks on data.
C++	32	0.814s	24.112s	12.3 bn/s	40.83x	No checks on data.
C++	63	0.721s	43.887s	13.9 bn/s	46.10x	No checks on data. Optimal run.
C++	64	0.733s	43.322s	13.6 bn/s	45.34x	No checks on data.