Except if the bit error happens in a pointer and you end up reading from the wrong part of memory. Or in your file system driver and you corrupt the file system. Yes, it will most likely be in the 500GB of float arrays, but why would you risk it?
Re:Best CPUs for Workstations: 2017
by iainmerrick on Thursday October 05, 2017 @11:09AM
(#15409773)
Exactly. I'm amazed at how blithely people dismiss the impact of bit errors. A random bitflip could break anything! That's terrifying.
Even if it's just in a float, it looks like a single bitflip could change FLOAT_MAX to NaN. That could hurt.
Re:Best CPUs for Workstations: 2017
by cjbillington on Thursday October 05, 2017 @11:36AM
(#15410009)
Yeah but you'll notice, and resume the simulation from a previous checkpoint or something. The risk is with errors that you don't notice, but that nonetheless change the outcome in a meaningful way. That seems to me to actually be a pretty small space of situations.
Sounds like we need a kernel module to introduce bit flips for bespoke errors.
The Linux EDAC (for ECC memory controllers) module allows injecting address-based errors (simulating the MC reporting a fault).
You can also use ptrace to corrupt the memory of a process more subtly.
See e.g. this paper by Walker and Betz from the San Diego Supercomputing Centre (paper from XSEDE '13) on the effect of ECC on GPUs for MD simulations: https://dl.acm.org/citation.cfm?id=2484774