Dead loop of counting the number of possible activations per refresh interval

[TheSilentDawn]

Hi,
I try to run Blacksmith on my machine with an i7-8700 CPU, Ubuntu 18 with Linux version 5.4.0-84-generic, and 1x16G Hynix/Hyundai HMA82GU6DJR8N-VK DRAM.
However, the program is stuck at the function DramAnalyzer::count_acts_per_trefi(). Based on my understanding, it's due to a hard-coded value 3.0 in

blacksmith/src/Memory/DramAnalyzer.cpp

Line 165 in 0d775c9

if ((acts.size()%200)==0 && compute_std(acts, running_sum, acts.size())<3.0) break;

Could you please explain why 3.0 is chosen and how to determine this value in other machines?

[its-luca]

I am not the author of blacksmith but also had some issues with this on some machines.

To my understanding this is a "gut instinct" boundary for the standard derivation of the measurements. The idea here is that we want to sample long enough for the sampled values to be fairly regular/stable.
To improve the timing accuracy you can try to pin blacksmith to an isolated CPU core.

Otherwise I would suggest to dump the gathered timings to see where things go wrong (i.e. if there are some heavy outliers). On some machines tweaking the timing function worked for me.

[TheSilentDawn]

Hi @its-luca ,
Thanks for your prompt reply.

Otherwise I would suggest to dump the gathered timings to see where things go wrong (i.e. if there are some heavy outliers). On some machines tweaking the timing function worked for me.

Could you please give a detail on how to tweak it? Just change the value 3.0 to other values based on the gathered timings?

[its-luca]

For me, changing the actual time measurement code a little improved the stability of the measurements on some machines. I'd have to look up specifics.
You can ofc. also increase the threshold for the std. However, if your measurement is too noisy it will be of little use. I cannot give you a specific upper boundary here.

Another error source could be the selection of the two addresses used in the experiment. They have be in the same bank and are sampled in

blacksmith/src/Memory/DramAnalyzer.cpp

Line 6 in 0d775c9

void DramAnalyzer::find_bank_conflicts() {

Did you adjust the timing threshold in

blacksmith/include/GlobalDefines.hpp

Line 50 in 0d775c9

#define THRESH 495 // worked best on DIMM 6

to your machine?

For debugging, I would again dump the timing values of this experiment, plot a histogram and check that it resembles Figure 3 in the drama paper (i.e. one large heap and a smaller one. The actual timing values might vary).