cluster computing

[jproffer]

Now using power via Euler bus.  Test board (top) has a heatsink on it also.  During stress testing (sysbench), it does get quite warm, but not hot. 

   

Test execution summary:
    total time:                          1290.0480s
    total number of events:              4000000
    total time taken by event execution: 5159.0812
    per-request statistics:
         min:                                  1.28ms
         avg:                                  1.29ms
         max:                                 11.99ms
         approx.  95 percentile:               1.29ms

Threads fairness:
    events (avg/stddev):           1000000.0000/1266.07
    execution time (avg/stddev):   1289.7703/0.01

[pineresearch]

Thank you for the stats. I got the whole wiring thing down now.



This was just about how you did it right? (Except I think you used a modded PC powersupply)

[jproffer]

Thank you for the stats. I got the whole wiring thing down now.



This was just about how you did it right? (Except I think you used a modded PC powersupply)

hmm if you're using the power supply I linked you to on Amazon, then that schematic is wrong.  The power supply has 3 (+) terminals and 3 (-) terminals.  The (-) terminals are the ground terminals for the +5v supply.

[tkaiser]

During stress testing (sysbench), it does get quite warm, but not hot. 

Sysbench is pretty useless. If you want to explore the thermal design of your cluster experiments (that are responsible for performance. On any modern SBC you get full performance only when thinking about thermal issues, cpufreq and dvfs operating points) better have a look at cpuminer: http://linux-sunxi.org/User:Tkaiser#Reli..._on_Pine64

[jproffer]

During stress testing (sysbench), it does get quite warm, but not hot. 

Sysbench is pretty useless. If you want to explore the thermal design of your cluster experiments (that are responsible for performance. On any modern SBC you get full performance only when thinking about thermal issues, cpufreq and dvfs operating points) better have a look at cpuminer: http://linux-sunxi.org/User:Tkaiser#Reli..._on_Pine64

I used cpuburn previously, it cycled through with no problems.   I'll try cpuminer next.

[tkaiser]

I'll try cpuminer next.

Well, cpuminer is nice since it not only creates significant load (unlike many other tools used for benchmarking) but also provides real performance numbers. And at the link above I provide a whole archive with tools and templates to simply measure/collect/graph the efficiency of performance (which is thermal) tuning. See the khash/s on the lower right corner:



Using RPi-Monitor you get also a history of measurements so you can easily compare. If I would setup a cluster I would immediately start to undervolt the individual cluster nodes (which requires testing for reliability and every board might behave different -- for this purpose Linpack is great) since this helps improving performance a lot due to throttling jumping in later.

Fun fact: as usual with benchmarking nearly all the times it goes wrong. Back when I had two Pine64+ here (I sent one of them to another Armbian dev, he received it yesterday and immediately gave it a go) I compared two different OS images by exchanging SD cards. One installation was approx. 5% faster. The only difference I found was one minor GCC update missing. And after updating GCC on the other installation performance was on par again.

The usual 'Casual benchmarking: you benchmark A, but actually measure B, and conclude you've measured C.'

[pineresearch]

Thank you for the stats. I got the whole wiring thing down now.



This was just about how you did it right? (Except I think you used a modded PC powersupply)

hmm if you're using the power supply I linked you to on Amazon, then that schematic is wrong.  The power supply has 3 (+) terminals and 3 (-) terminals.  The (-) terminals are the ground terminals for the +5v supply.

Thanks for the feedback.

 I bought one of off ebay instead as the shipping was cheaper to Europe. It looked like this. Each of those 3 +- terminals spit out 5V right? and the amp is just distributed equally across the if you want to use them all?

[jproffer]

I'll try cpuminer next.

Well, cpuminer is nice since it not only creates significant load (unlike many other tools used for benchmarking) but also provides real performance numbers. And at the link above I provide a whole archive with tools and templates to simply measure/collect/graph the efficiency of performance (which is thermal) tuning. See the khash/s on the lower right corner:



Using RPi-Monitor you get also a history of measurements so you can easily compare. If I would setup a cluster I would immediately start to undervolt the individual cluster nodes (which requires testing for reliability and every board might behave different -- for this purpose Linpack is great) since this helps improving performance a lot due to throttling jumping in later.

Fun fact: as usual with benchmarking nearly all the times it goes wrong. Back when I had two Pine64+ here (I sent one of them to another Armbian dev, he received it yesterday and immediately gave it a go) I compared two different OS images by exchanging SD cards. One installation was approx. 5% faster. The only difference I found was one minor GCC update missing. And after updating GCC on the other installation performance was on par again.

The usual 'Casual benchmarking: you benchmark A, but actually measure B, and conclude you've measured C.'

   
It's been running cpuminer for about 30-40 minutes so far

[tkaiser]

It's been running cpuminer for about 30-40 minutes so far

Hey, that looks pretty good! With your cooling approach you might be able to activate 1200 MHz and maybe slightly exceed it (1248 might work) when you follow the Hardware reliability tests outlined in linux-sunxi wiki.

And if you further tune you should use script/template provided in the aforementioned link to also monitor khash/s rate (at least when we did the dvfs/cpufreq finetuning back in March I used the khash/s rate in different scenarios -- no heatsink, heatsink, heatsink+fan -- as the tool of choice)

[jproffer]

 I bought one of off ebay instead as the shipping was cheaper to Europe. It looked like this. Each of those 3 +- terminals spit out 5V right? and the amp is just distributed equally across the if you want to use them all?

Usually it's all the same power bus, but you should check your power supply's documentation.  Some do tie each terminal to a separate regulator circuit.