4 ms·

Interesting, could you possibly explain how "halt behavior" indicates that ECC is fully functioning? I've just read a lot about ECC being compatible but not te

Interesting, could you possibly explain how "halt behavior" indicates that ECC is fully functioning? I've just read a lot about ECC being compatible but not tests explicitly showing it works.

ChuckMcM 9y ago
On my system if an uncorrectable error were to happen the system goes into a 'machine check' state (a halt). In theory, the BIOS supports 'chip kill' which is that it reboots with a bit set to not use that specific chip in the DIMM. I've never seen that happen on my desktop but in the data center when it happened a system that should have 128GB of memory would reboot and come back with less than the full amount (as expected by the DIMMs plugged in) reported by the BIOS.
- foota 9y ago
  Interesting, depending on the amount of RAM killed that seems like it could be a bit overkill for what could just be a cosmic ray flip. Though perhaps the rare frequency of cosmic ray flips makes that acceptable.
- fulafel 9y ago
  Chipkill is an IBM name for the feature that corrects multiple bit errors in DRAM, are you sure about the reboot thing or are there two similarly named ECC techs that do different things?
  - ChuckMcM 9y ago
    Intel discussion of Chipkill (https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/xeon-e7-family-ras-server-paper.pdf https://www.intel.com/content/dam/www/public/us/en/documents...). Servers in question were Supermicro "Jupiter" chassis with the Westmere processors, discussion on Supermicro of their implementation (https://www.supermicro.com/support/faqs/faq.cfm?faq=2642 https://www.supermicro.com/support/faqs/faq.cfm?faq=2642) When systems rebooted with less memory than the system configuration database said they should have most of the time there would be a multi-bit error detection, machine check, and 'memory update' in the IPMI buffer.
- kobeya 9y ago
  Why kill the DIMM? Cosmic ray bitflips don't harm the hardware, AFAIK.
  - SamReidHughes 9y ago
    It's usually not cosmic ray.
loeg 9y ago
The system must be configured to halt the system on detection of uncorrectable multibit errors -- which shows that the detection of uncorrectable errors works.

Interesting, could you possibly explain how "halt behavior" indicates that ECC is fully functioning? I've just read a lot about ECC being compatible but not tests explicitly showing it works.

ChuckMcM 9y ago

On my system if an uncorrectable error were to happen the system goes into a 'machine check' state (a halt). In theory, the BIOS supports 'chip kill' which is that it reboots with a bit set to not use that specific chip in the DIMM. I've never seen that happen on my desktop but in the data center when it happened a system that should have 128GB of memory would reboot and come back with less than the full amount (as expected by the DIMMs plugged in) reported by the BIOS.

foota 9y ago
Interesting, depending on the amount of RAM killed that seems like it could be a bit overkill for what could just be a cosmic ray flip. Though perhaps the rare frequency of cosmic ray flips makes that acceptable.
fulafel 9y ago
Chipkill is an IBM name for the feature that corrects multiple bit errors in DRAM, are you sure about the reboot thing or are there two similarly named ECC techs that do different things?
- ChuckMcM 9y ago
  Intel discussion of Chipkill (https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/xeon-e7-family-ras-server-paper.pdf https://www.intel.com/content/dam/www/public/us/en/documents...). Servers in question were Supermicro "Jupiter" chassis with the Westmere processors, discussion on Supermicro of their implementation (https://www.supermicro.com/support/faqs/faq.cfm?faq=2642 https://www.supermicro.com/support/faqs/faq.cfm?faq=2642) When systems rebooted with less memory than the system configuration database said they should have most of the time there would be a multi-bit error detection, machine check, and 'memory update' in the IPMI buffer.
kobeya 9y ago
Why kill the DIMM? Cosmic ray bitflips don't harm the hardware, AFAIK.
- SamReidHughes 9y ago
  It's usually not cosmic ray.

loeg 9y ago

The system must be configured to halt the system on detection of uncorrectable multibit errors -- which shows that the detection of uncorrectable errors works.