Otdfbt

I am looking for a quick and easy program to estimate FLOPS on my Linux system. I found HPL, but getting it compiled is proving to be irritating. All I need is a ballpark estimate of the FLOPS, without needing to spend a day researching benchmark packages and installing dependent software. Does any such program exist? Would it be sufficient to write a C program that multiples two floats in a loop?

The question is what do you mean by flops? If all you care about is how many of the simplest floating point operations per clock, it is probably 3x your clock speed, but that is about as meaningless as bogomips. Some floating point ops take a long time (divide, for starters), add and multiply are typically quick (one per fp unit per clock). The next issue is memory performance, there is a reason the last classic CRAY had 31 memory banks, ultimately CPU performance is limited by how fast you can read and write to memory, so what level of caching does your problem fit in? Linpack was a real benchmark once, now it fits in cache (L2 if not L1) and is more of a pure theoretical CPU benchmark. And of course, your SSE (etc) units can add floating point performance too.

What distro do you run?

This looked like a good pointer: http://linuxtoolkit.blogspot.com/2009/04/intel-optimized-linpack-benchmark-for.html

http://onemansjourneyintolinux.blogspot.com/2008/12/show-us-yer-flops.html

http://www.phoronix-test-suite.com/ might be an easier way to install a flops benchmark.

Still I do wonder why you care, what you are using it for? If you just want a meaningless number, your systems bogomips is still right there in dmesg.

apparently there's a "sysbench" benchmark package and command:

sudo apt-get install sysbench (or brew install sysbench OS X)

run it like this:

sysbench --test=cpu --cpu-max-prime=20000 --num-threads=2 run

output for comparisons:

 total time: 15.3047s

ref: http://www.midwesternmac.com/blogs/jeff-geerling/2013-vps-benchmarks-linode

For ballpark-estimates:

• Raspberry Pi 2: 299.93 * 10^6 FLOPS (source)




• Raspberry Pi 3: 462.07 * 10^6 FLOPS (source)



• 
  GTX Titan Black GPU: 5.1 * 10^12 FLOPS (source)


• 
  Sunway TaihuLight: 93 * 10^15 FLOPS (source, record holder of 2016)

Linpack

1. Download it (link)


2. Extract it


3. cd benchmarks_2017/linux/mkl/benchmarks/linpack


4. ./runme_xeon64


5. Wait for quite a while (more than 1 hour)

On a Thinkpad T460p (Intel i7-6700HQ CPU), it gives:

This is a SAMPLE run script for SMP LINPACK. Change it to reflect
the correct number of CPUs/threads, problem input files, etc..
./runme_xeon64: 33: [: -gt: unexpected operator
Mi 21. Dez 11:50:29 CET 2016
Intel(R) Optimized LINPACK Benchmark data

Current date/time: Wed Dec 21 11:50:29 2016

CPU frequency: 3.491 GHz
Number of CPUs: 1
Number of cores: 4
Number of threads: 4

Parameters are set to:

Number of tests: 15
Number of equations to solve (problem size) : 1000 2000 5000 10000 15000 18000 20000 22000 25000 26000 27000 30000 35000 40000 45000
Leading dimension of array : 1000 2000 5008 10000 15000 18008 20016 22008 25000 26000 27000 30000 35000 40000 45000
Number of trials to run : 4 2 2 2 2 2 2 2 2 2 1 1 1 1 1 
Data alignment value (in Kbytes) : 4 4 4 4 4 4 4 4 4 4 4 1 1 1 1 

Maximum memory requested that can be used=9800701024, at the size=35000

=================== Timing linear equation system solver ===================

Size LDA Align. Time(s) GFlops Residual Residual(norm) Check
1000 1000 4 0.014 46.5838 1.165068e-12 3.973181e-02 pass
1000 1000 4 0.010 64.7319 1.165068e-12 3.973181e-02 pass
1000 1000 4 0.009 77.3583 1.165068e-12 3.973181e-02 pass
1000 1000 4 0.010 67.0096 1.165068e-12 3.973181e-02 pass
2000 2000 4 0.064 83.6177 5.001027e-12 4.350281e-02 pass
2000 2000 4 0.063 84.5568 5.001027e-12 4.350281e-02 pass
5000 5008 4 0.709 117.6800 2.474679e-11 3.450740e-02 pass
5000 5008 4 0.699 119.2350 2.474679e-11 3.450740e-02 pass
10000 10000 4 4.895 136.2439 9.069137e-11 3.197870e-02 pass
10000 10000 4 4.904 135.9888 9.069137e-11 3.197870e-02 pass
15000 15000 4 17.260 130.3870 2.052533e-10 3.232773e-02 pass
15000 15000 4 18.159 123.9303 2.052533e-10 3.232773e-02 pass
18000 18008 4 31.091 125.0738 2.611497e-10 2.859910e-02 pass
18000 18008 4 31.869 122.0215 2.611497e-10 2.859910e-02 pass
20000 20016 4 44.877 118.8622 3.442628e-10 3.047480e-02 pass
20000 20016 4 44.646 119.4762 3.442628e-10 3.047480e-02 pass
22000 22008 4 57.918 122.5811 4.714135e-10 3.452918e-02 pass
22000 22008 4 57.171 124.1816 4.714135e-10 3.452918e-02 pass
25000 25000 4 86.259 120.7747 5.797896e-10 3.297056e-02 pass
25000 25000 4 83.721 124.4356 5.797896e-10 3.297056e-02 pass
26000 26000 4 97.420 120.2906 5.615238e-10 2.952660e-02 pass
26000 26000 4 96.061 121.9924 5.615238e-10 2.952660e-02 pass
27000 27000 4 109.479 119.8722 5.956148e-10 2.904520e-02 pass
30000 30000 1 315.697 57.0225 8.015488e-10 3.159714e-02 pass
35000 35000 1 2421.281 11.8061 1.161127e-09 3.370575e-02 pass

Performance Summary (GFlops)

Size LDA Align. Average Maximal
1000 1000 4 63.9209 77.3583 
2000 2000 4 84.0872 84.5568 
5000 5008 4 118.4575 119.2350
10000 10000 4 136.1164 136.2439
15000 15000 4 127.1586 130.3870
18000 18008 4 123.5477 125.0738
20000 20016 4 119.1692 119.4762
22000 22008 4 123.3813 124.1816
25000 25000 4 122.6052 124.4356
26000 26000 4 121.1415 121.9924
27000 27000 4 119.8722 119.8722
30000 30000 1 57.0225 57.0225 
35000 35000 1 11.8061 11.8061 

Residual checks PASSED

End of tests

Done: Mi 21. Dez 12:58:23 CET 2016

One benchmark that has been traditionally used to measure FLOPS is Linpack. Another common FLOPS benchmark is Whetstone.

More reading:
The Wikipedia "FLOPS" entry,
Whetstone entry,
Linpack entry

I highly recommend the ready-to-run linpack build from Intel:
http://software.intel.com/en-us/articles/intel-math-kernel-library-linpack-download/

As you mention cluster, we have used the the HPCC suite. It takes a bit of effort to setup and tune, but in our case the point wasn't bragging per se, it was part of the acceptance criteria for the cluster; some performance benchmarking is IMHO vital to ensure that the hardware works as advertised, everything is cabled together correctly etc.

Now if you just want a theoretical peak FLOPS number, that one is easy. Just check out some article about the CPU (say, on realworldtech.com or somesuch) to get info on how many DP FLOPS a CPU core can do per clock cycle (with current x86 CPU's that's typically 4). Then the total peak FLOPS is just

number of cores * FLOPS/cycle * frequency

Then for a cluster with IB network you should be able to hit around 80% of the peak FLOPS on HPL (which BTW is one of the benchmarks in HPCC).