Hardware Issues

Introduction

Conditions under which the hardware of the cluster are kept.

Service

• The Supermicro servers were supplied by Viglen, which was bought by, and is now known as XMA.
• Serives terms are "Next Day" which is neither a high, nor low service level.
• The duration of the warranty is 5 years and comes to an end on 4th Nov 2018.

Details

• Contact email is gold@xma.co.uk
• Contact number is 01727 201 850
• Favourite contact: Doshunn. Least favourite contact: Bradley.
• Local engineer is Graeme Akers, tel. 07970 444162, but he can only be contacted via central support.

Submitting incidents to XMA

There is no particular account number for our warranty, but XMA work mainly with product serial numbers so mentioning the serial number of any of the products we've bought from them will be enough for the agent to call up our contract. For example

• Serial HPC-2367949 is marvin itself, the front-end or head-node
• Serial HPC-2367837-b is node2, whose entire motherboard was replaced in January 2016.
• Serial HPC- 2367950 is node10.

Incidents

Repeated memory warnings on marvin front-end

Because these self correct (due to the RAM being expensive ECC RAM) it look as if these warnings can be safely ignored.

Unfortunately, all the users see them and are quite mystified by them.

Here is one example from 16/12/2016

Message from syslogd@marvin at Dec 16 13:03:27 ...
 kernel:[Hardware Error]: MC4 Error (node 0): DRAM ECC error detected on the NB.

Message from syslogd@marvin at Dec 16 13:03:27 ...
 kernel:[Hardware Error]: Error Status: Corrected error, no action required.

Message from syslogd@marvin at Dec 16 13:03:27 ...
 kernel:[Hardware Error]: CPU:0 (15:2:0) MC4_STATUS[-|CE|MiscV|-|AddrV|-|-|CECC]: 0x9c00400001080a13

Message from syslogd@marvin at Dec 16 13:03:27 ...
 kernel:[Hardware Error]: MC4_ADDR: 0x0000000c00b0ef80

Message from syslogd@marvin at Dec 16 13:03:27 ...
 kernel:[Hardware Error]: cache level: L3/GEN, mem/io: MEM, mem-tx: RD, part-proc: RES (no timeout)

MC4 appears to be the offending DIMM. Here is another from 13 Dec:

Message from syslogd@marvin at Dec 13 15:48:27 ...
kernel:[Hardware Error]: MC4 Error (node 0): DRAM ECC error detected on the NB.

Message from syslogd@marvin at Dec 13 15:48:27 ...
kernel:[Hardware Error]: Error Status: Corrected error, no action required.

Message from syslogd@marvin at Dec 13 15:48:27 ...
kernel:[Hardware Error]: CPU:0 (15:2:0) MC4_STATUS[-|CE|MiscV|-|AddrV|-|-|CECC]: 0x9c00400001080a13

Message from syslogd@marvin at Dec 13 15:48:27 ...
kernel:[Hardware Error]: MC4_ADDR: 0x0000000c00b0ef80

Message from syslogd@marvin at Dec 13 15:48:27 ...
kernel:[Hardware Error]: cache level: L3/GEN, mem/io: MEM, mem-tx: RD, part-proc: RES (no timeout)

There is a tool called edac-util whihc allows you to see errors, but it doesn't correlate with MC4, claiming in fact that it is MC0 that is correcting errors. Nevertheless when a warning comes in about MC4, the MC0:csrow0:ch0 DIMM coordinate gets incremented ... so the warnings do correlate with edac-util.

Back on 3 Nov there was also:

Message from syslogd@marvin at Nov  3 19:15:33 ...
kernel:[Hardware Error]: MC4 Error (node 0): DRAM ECC error detected on the NB.

Message from syslogd@marvin at Nov  3 19:15:33 ...
kernel:[Hardware Error]: Error Status: Corrected error, no action required.

Message from syslogd@marvin at Nov  3 19:15:33 ...
kernel:[Hardware Error]: CPU:0 (15:2:0) MC4_STATUS[-|CE|MiscV|-|AddrV|-|-|CECC]: 0x9c00400001080813

Message from syslogd@marvin at Nov  3 19:15:33 ...
kernel:[Hardware Error]: MC4_ADDR: 0x0000000c00b0ef80

Message from syslogd@marvin at Nov  3 19:15:33 ...
kernel:[Hardware Error]: cache level: L3/GEN, mem/io: MEM, mem-tx: RD, part-proc: SRC (no timeout)

Looking over emails, this also occurred Aug19 reported by Luke and I have some other error reports. It now appears that the errors are the exact same all the time, referring to this same MC4_ADDR.

It seems fairly clear there is only one source for these errors, so it might be an idea to find "MC4" or "MC0:CSRW0:CH0". dmidecode is a linux command that give quite useful information on the DIMM banks: For example:

Handle 0x005B, DMI type 17, 28 bytes
Memory Device
       Array Handle: 0x001B
       Error Information Handle: Not Provided
       Total Width: 72 bits
       Data Width: 64 bits
       Size: 16384 MB
       Form Factor: DIMM
       Set: None
       Locator: P4_4B
       Bank Locator: 31
       Type: DDR3
       Type Detail: Synchronous Registered (Buffered)
       Speed: 1600 MHz
       Manufacturer: Hyundai    
       Serial Number: 38589658
       Asset Tag:  
       Part Number: HMT42GR7MFR4C-PB  
       Rank: 2

This is the last record, so marvin has 32 such banks, which most likely correspond to separate DIMMS. So the hunch right now is that one of these has been requiring all this correction. The dmidecode schema is as follows:

P{1,2,3,4}_{1,2,3,4}{A,B}

Correspnding to banks 0 to 31 of 16G each which is 512G RAM, correct.

The syslogd warnings do not show one particular line, but this can be found in /var/log/messages:

EDAC MC0: CE page 0xc00b0e, offset 0xf80, grain 0, syndrome 0x100, row 0, channel 0, label "": amd64_edac

It appears now that the MC4 name in the warnings may be a bit of a red herring and that the problem is MC0. The fibrevillage link below shows you can append the CE page and offset address to give the address like so

0xc00b0ef80

This appears quite small and so is somewhat disconcerting, but what seems likely is that it's the veryfirst memory module that is causing the problems.

dmidecode -t 20

Will list out the type 20 devices and allow us see the addresses for the modules.

Links

• fibrevillage edac-util and dmidecode