Relating MemTest86's ECC error "Channel-Slot" format to actual system hardware?

[ZPrime]

PassMark team and other internet people:

I'm working for a nonprofit and trying to test some DDR4 ECC memory in a non-production server before moving it into live systems. I have a total of 18 memory sticks I need to test prior to installing them in 3 separate production servers (each prod server takes 6 sticks, hence 18 total). We don't have money to buy a licensed copy, but I appreciate the existence of the free MT86!

The test rig is a less-common model of Dell server meant to be used in networking scenarios and other non-standard deployments, a PowerEdge "XR11".
It uses a 3rd-gen Xeon Scalable CPU (in our case, a single Xeon Silver 4314). It has 8 total memory slots, arranged 4 on each side of the socket.

On the test server, the 8 slots are labeled "A1" through "A8", and the installation manual says that RAM should be installed in numeric order.


The Dell installation/service manual also has this table in it -
"Memory channels are organized as follows:"

Channel A	Channel B	Channel C	Channel D	Channel E	Channel F	Channel G	Channel H
Slot A1	Slot A5	Slot A3	Slot A7	Slot A2	Slot A6	Slot A4	Slot A8

Running a test with all 8 slots full, I saw some ECC Errors, which the MT86 output labeled as "Channel-Slot: 7-0" on several tests, plus we saw some failures on "Channel-Slot: 0-0" for Test 13.

If I convert the "Channel Letter" system to a 0-based number system, this would imply that "0-0" is Channel A / Slot A1, and "7-0" is Channel H / Slot A8, right?

I just removed all memory from the system, and relocated the stick from A8 over to A4 (so Channel H to Channel G), and the stick from A1 to A3 (Channel A to Channel C). I then took two other sticks that were error-free but had only been tested for ~8 hours (rather than a full 4 passes), and put those into A1 and A2, just to give them a more thorough test.

I'm already seeing some ECC errors in this test, but this time from "Channel-Slot: 6-0," which I think would correspond to Channel G / slot A4, and an error from "Channel-Slot: 2-0," which I believe is Channel C / Slot A3.

Am I "decoding" the MemTest86 designations correctly here? I did save a report from the initial test, but the USB stick with the report is still in the server running the second test. I can't grab the HTML report and attach it until this round of testing completes (at which point I should have two different reports).

[David (PassMark)]

Memtest86 just reports the numbers from the CPU's memory controller. We can't be sure about how Dell numbered the slots and if they line up with the Xeon Silver 4314 view of the world. But in the absence of support from Dell, you are doing the right thing by moving the sticks and checking if the errors move with the sticks. But with multiple suspect sticks at the same time it gets harder to work out what is going on.

Another approach would be to just test 1 stick at a time, then you don't need to guess about the mapping. Then once you know the mapping, go back to testing 8 at a time.

[ZPrime]

David, thanks for the reply. Based on Dell's own published table combined with what I observed from moving sticks, I'm pretty sure that "Channels A-H" correspond to "0-7" in MT86's reporting. The system only has one slot per memory channel.

Good point on testing a single stick at a time. The test design was structured around when I had physical access to the test system and how long I knew it would take to test 512GB thoroughly vs only 128GB. So I did 8 sticks, then 2 on their own on a shortened timeline, and then another 8.

I didn't get any failures until the final set of 8 sticks, but I didn't give that "middle pair" a full workout either. I was trying to kill two birds with one stone by doing more testing of the middle pair, and also double-check that the errors from the other two sticks were not just flukes, hence trying to test 4 now, but shuffling things around (to try to insure the errors "moved" with the bad sticks).

I'll probably go back and give each of the bad sticks one more partial run alone to be absolutely certain.