[3dem] weird server issues, suggestions/advice would be helpful!

[Ludtke, Steven J.]

Hi Liz,
with a machine like this, I suspect you purchased from a vendor, rather than building it yourself?  If so, this seems an issue for vendor support.  Other things you omitted:

- power supply model and ratings
- how many of the Xeon Gold processors?
- have you run a memory test

power is the most likely culprit...

--------------------------------------------------------------------------------------
Steven Ludtke, Ph.D. <sludtke at bcm.edu<mailto:sludtke at bcm.edu>>                      Baylor College of Medicine
Charles C. Bell Jr., Professor of Structural Biology
Dept. of Biochemistry and Molecular Biology                      (www.bcm.edu/biochem<http://www.bcm.edu/biochem>)
Academic Director, CryoEM Core                                        (cryoem.bcm.edu<http://cryoem.bcm.edu>)
Co-Director CIBR Center                                    (www.bcm.edu/research/cibr<http://www.bcm.edu/research/cibr>)



On Mar 23, 2020, at 5:25 PM, Liz Kellogg <lizkellogg at GMAIL.COM<mailto:lizkellogg at GMAIL.COM>> wrote:

Hi 3dem-ers,

I hope everyone is safely at home.  I have a non COVID-19 related problem that I hope others can help advise on.

I bought a 8-GPU server last April that would be used by my lab for image processing work. Everything seemed fine initially, however once I started getting more users the server became noticeably unstable around December and started randomly rebooting itself. It was happening so often that at its worst we couldn’t get through a single refinement job without a reboot.  Here are some technical details and hints at what could be going wrong:

Configuration of the server:
TYAN Thunder HX FT77D-B7109 8GPU 2P 4x3
Intel Xeon Gold 6138 20C 2.0-3.7 GHz
384 GB DDR4 2400/2666 ECC/REG (12x32GB)
SamSung 480GB 883 DCT SSD x 2
Seagate 12TB SAS x 16
GeForce RTX-2080Ti 11 GB x 8

The most noticeable errors we see when the server is up are the GPU devices becoming undetectable, along the lines of:

$ nvidia-smi
Unable to determine the device handle for GPU 0000:B1:00.0: GPU is lost.  Reboot the system to recover this GPU

Or

$ nvidia-smi
No devices were found

Replacing the GPUs did not seem to help which we did back in January, we are back to the same issues.
We also tried updating the GPU drivers to NVIDIA-SMI 440.33.01    Driver Version: 440.33.01 (before they were 410.48)
However, we experience pretty much the same behavior before and after the driver update.

Since we have updated the drivers, I doubt that’s a driver issue. Although it could be  a PCI bus issue, doesn’t seem likely to me because each of the 8 cards tend to go down randomly (during one strange episode, they were flickering on and off). My gut feeling is that there is either a power issue where the system’s power was not dimensioned properly (though looking at the chassis specs this seems unlikely as well), or a cooling issue. I am planning on monitoring the GPU temperature (I wrote a bash script using nvidia-smi -q) under heavy load and see if the current temp exceeds the maximum temp of each GPU.

Any idea of what would be going on? I think I have pretty standard server config.. has anyone experienced similar problems? Anyone that has configurations that work well for you, would you mind sharing your specs and your NVIDIA driver versions? Even if it's exactly the same specs that would help. Any non-standard steps to configure the machine or the drivers? I am mystified as to why we are experiencing these issues.. and doesn’t help that we’re all working from home at the moment :*(

Thanks everyone, stay safe.

Best wishes,

Liz

Elizabeth H. Kellogg, Ph.D.
Assistant Professor, Cornell University
Molecular Biology and Genetics

_______________________________________________
3dem mailing list
3dem at ncmir.ucsd.edu<mailto:3dem at ncmir.ucsd.edu>
https://urldefense.proofpoint.com/v2/url?u=https-3A__mail.ncmir.ucsd.edu_mailman_listinfo_3dem&d=DwICAg&c=ZQs-KZ8oxEw0p81sqgiaRA&r=Dk5VoQQ-wINYVssLMZihyC5Dj_sWYKxCyKz9E4Lp3gc&m=I-DCLW8-a8G9CUwNPojXeXKScOa76l1CszApIJVedsI&s=Xhi1X3aXYltnUN2kig-0Dgg45iF3QszB2XwpyImF4v0&e=

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.ncmir.ucsd.edu/pipermail/3dem/attachments/20200323/a75abfd0/attachment-0001.html>