WhyAmIgettingMemoryErrors

From EdacWiki
Jump to: navigation, search

To help understand why you are seeing memory errors, please have a look at HowMemoryEdacHardwareWorks.

You may well have been experiencing these errors for a while, it's just that nothing was checking them until you enabled the EDAC module. Note that your system is probably experiencing data corruption (if you are getting UEs - uncorrectable errors), so you should really check this out (this is why EDAC is set to `panic()` on UEs by default).

The reason that you are seeing problems is very likely to be one of:

  • Your RAM is bad.
  • Your Motherboard is bad.
  • Your CPU is bad (for CPUs which have the memory controller built into the CPU core, such at the AMD Opteron/Athlon-64).
  • The connection between your motherboard and your CPU, or memory module is bad.
  • Some of your hardware is being operated outside of its design specification, such as:
    • Things are being run too hot.
    • Timings are being violated (e.g. running memory too fast, or bad DRAM clock generation).
    • Supply voltages to the critical compontents are too high/low (this may even happend very briefly, as a supply "spike", or "droop").
  • You have seen one or more "Single Event Upsets" - see SoftErrors.
  • Memory ECC check bits are not properly initialised by BIOS prior to Linux boot. See Uninitialized ECC bits
  • The EDAC module is buggy.
  • Memory loading is exceeded.
  • The powersupply is insufficient.

[edit] So Which One Is It Then?

Good question. Time to try some things:

[edit] Symptoms

Here are the most likely symptoms.

Problem Error Addresses Error Slot or Row Error Frequency
Bad Memory Module(s) Single/Few Probably only 1 May vary if part is marginally out of spec
Bad Motherboard Probably many Maybe 1, maybe many  ?
Bad Connection Probably many 1 (bad mem), prob all (bad CPU)  ?
Temp out of spec Probably few Maybe one, if different mem mfrs/parts Usually higher, with higher temp
Timings out of spec Probably few Maybe one, if different mem mfrs/parts Usually higher, with higher temp
Voltages out of spec Probably random Maybe one, if different mem mfrs/parts Usually higher at higher system load
Bad BIOS check bit init Probably random Probably all High/very high (stops after a while for systems with background scrub)
Single event upsets Random Varies with effective "cross-section" of part Rare - more common with some part designs, and at high altitude etc.

[edit] Things to try to isolate the problem

General:

  • Get a second opinion e.g. from [[1]] or [[2]] - note that you should be sure that either:
    • The memory testing software knows how to disable ECC on your system, or
    • You have disabled ECC before running memory tester (note that memtest86 currently displays "ECC: No" on chipsets which have ECC, but which it doesn't know about!).
  • This may not catch problems like power-supply related problems, which don't occur when the memory tester is running.
  • Use a system stress tester such as "burnbx" from [[3]].
  • Put your system under stress by (e.g.) running a parallelised Linux kernel build, whilst doing some heavy 3D graphics display, and a lot of disk I/O.

Suspected bad module:

  • Remove Module.
  • Move Module to different slot (do errors move with module).
  • Move Module to different machine.
  • See "suspected temp out of spec".
  • See "suspected timings out of spec".
  • See "voltages out of spec".
  • Clean connections.
  • Check Memory Loading
    • Some memory controllers can only support so many 'ranks' of memory at a given speed.
    For example, Opterons/Athlon64s can support only 4 ranks of 2 GB at PC3200. 
    See http://www.valueram.com/memoryranks/default.asp for definitions.

Suspected bad motherboard:

  • Check motherboard docs for memory module compatability.
  • Move modules to different slots.
  • Clean connections.
  • Upgrade BIOS.
  • Select BIOS "fail-safe defaults", or equivalent change settings from there to isolate cause.

Suspected bad connection:

Suspected temp out of spec:

  • Measure temp, compare to published specs:
    • Use internal machine sensors (motherboard, hard drive etc.) if possible.
    • Use a temperature probe or infra-red thermometer.
  • Check airflow.
  • De-dust.
  • Lower temp:
    • Lower room temp.
    • Increase cooling.
    • Improve airflow (tidy cables etc.).

Suspected timings out of spec:

  • Try different BIOS version.
  • Set pessimistic memory timings in BIOS.
  • Compare memory controller timings to DIMM specs, using decode-dimms.pl from the Linux i2c project.
  • Try disabling "spread spectrum" in the BIOS (easy if available), or by using an i2c driver for your board's clock generator (hard).

Suspected voltage out of spec:

  • Check PSU specs vs. total demand of system components.
  • Swap power supply with another machine.
  • Fit voltage regulator/spike suppressor to machine power supply.

Suspected single event upsets:

  • Fit less susceptible components
  • Move to a lower altitude, or area with lower cosmic radiation.
  • Move your data centre underground.
  • Improve error-reporting utilities to ignore them.

Suspected bad check-bit init:

  • Upgrade BIOS.
  • Don't enable BIOS "quick boot".
  • Don't manually skip BIOS memory check.

Suspected insufficient powersupply:

  • Try detaching some devices that are hardly use. Start with USB devices.
  • If the problems stop, either structurally reduce the devices, or get a higher capacity powersupply.
  • Use a DC current clamp (pref one with peak/inrush measurement function) to check over-capacity at a particular voltage.
  • This is closely related to voltage out of spec. That can also be caused by just a broken supply.


Personal tools