[isabelle-dev] [158c513a39f5] JVM crash
eberlm at in.tum.de
Mon Sep 4 15:12:40 CEST 2017
Alas, it would appear I have spoken too soon!
I experienced a strange build failure with RC1 yesterday and, fearing
the worst, did my experiment from a few weeks ago again, with the
– building "Pure" fails in around 6 % of the cases
– this does not change even after a cold reboot
– switching SMT off seems to make the problem go away entirely
– switching SMT on makes it reappear
Sounds very much like this might well be the Ryzen bug. AMD has started
replacing affected CPUs, so I shall enquire about that and see what happens.
On 20/08/17 14:31, Manuel Eberl wrote:
> Addendum: I did a system upgrade (including a major kernel upgrade)
> around the same time when I first noticed the problem and I don't think
> I rebooted afterwards, so one very plausible explanation is that Linux
> introduced some workarounds/bug fixes in the kernel within the last few
> weeks that solved whatever issue I was having, and, of course, it took a
> reboot for it to kick in.
> On 20/08/17 14:14, Manuel Eberl wrote:
>> Okay I have no idea what is going on anymore.
>> I know have the following new data points:
>> – 2500 iterations on my Intel Core i7 laptop. No failures.
>> So it must be the Ryzen issue after all, I thought. I recalled that some
>> people said the problem was less pronounced with SMT disabled, so I
>> disabled it.
>> – 100 iterations on Ryzen 1800X after a reboot with SMT disabled. No
>> – 100 iterations on Ryzen 1800X after another reboot with SMT enabled
>> again. No failures.
>> So it seems the problem went away as mysteriously as it appeared and it
>> probably has something to do with the hardware or software constellation
>> on my computer. Or perhaps I should check my flat for radiation sources.
>> On 20/08/17 11:24, Lars Hupel wrote:
>>>> Lars, maybe you can run the same test on your machine and see what
>>>> happens there.
>>> I did, and nothing happened for about 100 iterations. I have a Core
>>> i7-2600. OS is otherwise identical to Manuel (Arch Linux).
>>>> As for Scala, could a problem in the Scala compiler really lead to the
>>>> JVM segfaulting? I would have thought if the JVM segfaults, that's a bug
>>>> in the JVM. (unless it's a hardware-related issue, of course)
>>> I've seen it happening, but it is very rare. Still, the coincidence of
>>> crashes during compilation could be explained by random chance (even if
>>> very unlikely). A quick look over Scala's issue tracker reveals no
>>> documented JVM segfaults after 2011.
More information about the isabelle-dev