Reproducible crash in DrizzleIntegration

[Juan Conejero]

Hi John,

No, this problem has not been reported before. However, I can easily reproduce it on my MBP placed directly on a table and allowing the process to use all available processor cores.

I have just completed a new test. I have integrated the 182 images again without problems after limiting the maximum number of processor cores to 4 (Edit > Global Preferences > Parallel Processing and Threads > Maximum number of processors used = 4, then Apply Global or F6). With just 4 processor cores, the task has required 1 hour and 18 minutes to complete. With 12 cores it required 52 minutes. This is a 50% time increase. During the drizzle integration, the machine has not perceptibly heated up and the fan has operated audibly only at times. I have placed a small wooden wedge of one centimeter in the back to facilitate ventilation. I'll repeat the same test with 6 processor cores (which is the number of physical cores in my machine).

[johnpane]

Thank you for running these tests! I will be very curious what you get with 6 cores. Most of the performance penalty should be recovered but maybe specifying 6 (instead of the default of 12) cores will reduce the load enough to let the processor run slightly cooler.

[johnpane]

Juan,

I just ran DrizzleIntegration with number of processor cores set to 6. The integration completed successfully and I did not do anything to enhance cooling. The computer was just sitting directly on my desk.

I noticed that the CPU frequency was actually higher than it was yesterday when I tried running this with a book propping up the back side. Yesterday, the CPU frequency peaked at about 3.08 GHz for each subframe, then throttled back to about 2.80 GHz. Today, it peaked around 3.55 GHz and throttled back to slightly over 3.00 GHz. In  both cases, the temperature peaked at about 100 C each subframe, but today it settled lower once throttling took effect.

The integration completed in 1:08:06, for a rate of 22.45 seconds per frame. I did not formally measure the rate when I had cores set to 12, but had informally estimated it to be about 20 seconds. Therefore, there was not a tremendous speed hit from decreasing cores from 12 to 6.

I am guessing that when the computation is cpu-bound, using hyperthreading actually increases the thermal load due to the necessity to swap back and forth between the two processes per core. On the other hand, if the computation has an I/O bottleneck this might not be the case. It is possible the performance hit from reducing the cores by half will be seen in other processes that may not be so cpu-bound.

John

[pfile]

100C! generally commercial silicon (non-milspec) would be designed for the slow-slow corner to still meet timing at a junction temperature of 85C. obviously there is always some headroom since the silicon is probably more typical than slow-slow, and intel may not follow the same methodology, but i'm not surprised to hear that some path failed @ 100C.

sounds like apple may still have some bugs in the thermal management stuff.

rob

[Niall Saunders]

Hi folks,

I gave up trying to rely on laptop machines for any serious kind of number-crunching - in my experience they have all failed (eventually) due to thermal issues.

Deskop machones - with planty pf internal air space and lots of forced ventilation have been what I have been relying on now for almost a decade. Laptops? A total waste of money, and energy, in my opinion - unless your needs are of the simplest variety.

As to raising a help ticket with Apple - I'd love to hear how you get om with that (you are probably 100 times more likely to get a result from Apple as you would from Microsoft - but 100 times 0 is stil 0  )

[bvalente]

i have an 18 core desktop the exhibits the same drizzle crash, but not on all drizzle integrations. I'm going to try to throttle to 6 cores and see what happens


Brian

[bvalente]

UPDATE: reducing PI to 6 cores seems to correct it here too


Brian

[niccoc1603]

Hi,
I have the same problem, I have crashes with LocalNormalization and Drizzle Integration, 150 frames

I am running PI on an i7 8700k with 12 logical processors.
It's a desktop PC with water cooling, max temp is 80°C at full load

One thing I have noticed when I launche LN the console shows "13 threads" instead of the standard 12

I will try limiting the threads to 10

I can share my dataset if needed

[bvalente]

you might try using process lasso, which is an application that enables you to easily change CPU assignments, etc.

i should have said 6 logical processors, but i think it's really 3 cores, 6 threads

still works every time, although it's much slower

Brian

[niccoc1603]

Limited to 10 cores, console shows 12 working threads, crashing


Limited to 8 cores, console shows 10 working threads, it works through the end


 I believe some optimizations are needed , hopefully I didn't buy this PC to run it at reduced potential...

[bvalente]

I understand believe me -

I built a computer just to do pixinsight, it's the #1 fastest on Pixinsight Benchmark PI Core 01.08.06.1457 (x64) (105)

Sometimes it requires a little patience at least we have a workaround for now

Brian

[pfile]

have you tried prime95 to see if that crashes your machine as well? PI is very taxing on a computer's resources.

how about memtest86+ ? maybe you have a bad dimm or something?

rob

[niccoc1603]

have you tried prime95 to see if that crashes your machine as well? PI is very taxing on a computer's resources.

how about memtest86+ ? maybe you have a bad dimm or something?

rob

Yes, and OCCP as well, I have no crashes

[niccoc1603]

This is the dataset if anyone wants to test it:

https://www.dropbox.com/s/ku2e0chmfux52hr/LNORM%20CRASH.rar?dl=0

[johnpane]

This case is interesting because it generalizes the problem beyond MacOS, laptops, and high CPU temperatures. It begins to suggest there may be an issue with the i7 processor itself.