Hello All,
I am running Fedora 21 x86_64, on a Haswell refresh i7-4790, Asus Z97M-plus MB.
I often build a number of packages from source(regular test builds from repos), and have been getting what appear to be random errors, since moving to this hardware. I never saw this problem on my old hardware running Fedora 20, building the same packages.
The errors are always of the form...

"internal compiler error: Segmentation fault
Please submit a full bug report,
with preprocessed source if appropriate.
The bug is not reproducible, so it is likely a hardware or OS problem."

and usually (but not always) show up when using multiple CPUs.
When this error occurs, I can start from scratch (ie clean build dir) and the project may succeed or may fail again, at a different place. If I persist I can always get a successful build.
Generally, if I run the build process using only one CPU, it will succeed.
I suspected a memory problem, and ran a memory test which had 100% success.
All of the projects use cmake.
I am using cmake-3.0.2-2.fc21.x86_64, make-4.0-3.fc21.x86_64, and gcc-4.9.2-1.fc21.x86_64.

Submitting a bug report is a bit problematic as these are large projects, and due to the random nature of the error and the lack of information, does seem like a bit of wild goose chase.

I would really like to get a better feel for the possible causes before going any further.

Any suggestions?

Cheers,
Terry

Hi,

sounds like bad ram to me too. Which memory test did you run?

Evo2.

I ran memtest86+ 5.1. I think it ran all the test suite.

Cheers,
Terry

I ran the memory tests again.
I ran all tests, parallel (all CPUs).
The tests took about 2.5 h (16GB), and the results are not at all clear.
On my system memtest86+ presents with a dreadful interface with some of it's display off screen, and I see no way of altering this.
When the tests were completed there were errors, but due to the screen display, it is unclear where they had occurred.
I have 2 x 8GB memory sticks, and it would be useful if memtest86+ could tell me which of those had errors.
Does memtest86+ also test CPU cache memory? From what I saw of the report it isn't clear to me whether the errors occurred in cache or RAM.
Any advice on how to configure the memtest86+ display, or other methods of testing that are less problematic?

Cheers,
Terry

Hi,

sorry I'm no expert on memtest86+, but a sure fire way to figure out if it is one stick or the other is to just put only one in and then run the test etc.

Evo2.

Errors in cpu or cpu cache is very unlikely. There are to ram testers. One memtest86 and other memtest86+. Try both.

Yes, that is a possible approach, although I may fall foul of the rotten display...ie the results I need to see being off screen.
I'll try this approach and see how I go.

Cheers,
Terry

I was under the impression that memtest86 was proprietary.
I'll have another look.

Cheers,
Terry

Well memtest86 was initially open source. Then it changed closed source with version 5 as I remember. Though source code for version 4.3 or perhaps 4.7 is available.

Though they provide a free binary of memtest86.

The plot thickens.
The tests I have done to date have actually been with memtest86, not sure how I got that mixed up.
I got hold of memtest86+ and tried that. It gives a much better display, everything on screen and readable.
My MB has four slots, L to R they are labelled A1, A2, B1, B2.
The advice is for one memory stick, use slot A2. For two sticks, use A2, B2.
I have tested each of my two 8GB sticks individually in slot A2 and neither give any errors.
When I have both sticks in slots A2, B2, I get about 260 errors.
So, what might be going on here? Could the problem be related to slot B2 in the MB?
I need some means of determining if the problem is memory or MB.
What consequences, if any, of fitting the sticks in slots A1, B1, to see if any errors detected with that setup?

Cheers,
Terry

...and thickens again!
Pondering on the results I have reported, I thought it wise to have a look at the BIOS settings.
The memory sticks were purchased as as "16 GB kit 2400 (2x8GB)", and I recall that whilst the BIOS detected the ram as 16GB 2400, it also reported the ram as 1333 MHz.
In the BIOS the memory frequency was set at 2400, so I reset to "auto".
Another run through with memtest86+ with both sticks installed reports no errors. probably should let it run multiple runs but I need the system for work.
So, the problem may be solved, I'll have to see if the unreproducable errors crop again when compiling. Here's hoping.

Cheers,
Terry

Hi,

excellent, sounds like you may have solved it. Let us know if you see any more problems.

Evo2.

So it's probably a overclocking issue. Setting it to auto in BIOS solved it.

Can you see at what speed it's running in memtest.

Overclocking can damage cpus or ram.

Overclocking also affects stability of rams. So errors in operation of ram may be produced.

If I recall correctly memtest reported the ram to be running at 2400 when the BIOS was set for 2400, and at 1333 when set to 'auto".
It hadn't been set to be overclocked as the ram was specified, as I said, as 2400, not 1333, but I just dredged up Kingston's specs for this ram, and whilst tested to run at 2400, 1.65v, are "...programmed to run at 1333 1.5v", so suspect that the MB doesn't like running the ram at 1.65v.

Cheers,
Terry