New Hardware, Who Dis?

After my last benchmarking post, I alluded to making some hardware upgrades. I completely shut down one of the older Dx0q nodes, but this still left me with one slower node and two quicker Wyse 5070 nodes. So I started watching eBay for models recommended on the ServeTheHome TinyMiniMicro and thin client forum threads, and ended up getting an HP T640 with 8GB of RAM for $110 with minor cosmetic damage. I swapped out the 16GB eMMC card for a 256GB NVME drive, and installed Debian Linux, because I haven’t been happy with Clear Linux’s Kubernetes update speed.

There was a bit of figuring out what software to get installed — I decided to try out containerd rather than cri-o for my runtime, but ended up going back when it turned out that I was getting SIGWINCH (window size changed) signals sent to the apache process that this WordPress sites runs under. SIGWINCH is used by Apache to signal “please gracefully shutdown”, which meant that WordPress never really got a chance to run. It’s mentioned in this issue, but I couldn’t find a resolution, and I wanted to actually get the hardware running.

The new hardware is hot!

The first time I tried benchmarking was while I was on the road at Kubecon, and I got the benchmark run started and walked away for a few days. When I came back, it seemed like the node was offline, and the job had errored. Uh-oh.

Getting home, I eventually got a chance to sit down with my handy display adapter and keyboard. No video, no reaction to keyboard. So I rebooted the unit, and saw a netboot (PXE) prompt. Huh, that’s odd. Rebooting to BIOS showed no attached storage devices. This was definitely going to require more work, so I unplugged the unit, pulled it out of the rack, and took off the lid. Everything inside was very toasty! I let things cool down a bit, and then plugged the ethernet, power, and keyboard/display back in. At this point, I discovered that there was some sort of secure boot warning that I needed to disable, but the hard drive was back (I also remembered to reduce the integrated graphics set-aside from 2GB to 256MB, which is as low as it goes).

I decided to re-run the benchmarks with the case off, since there’s no fan inside the unit to assist with airflow.

But, more importantly, did we get a charge?

Installing lm-sensors package and running it under watch, I saw 70C temperatures on both the CPU and the NVME card:

nvme-pci-0100
Adapter: PCI adapter
Composite:    +57.9°C  (low  =  -0.1°C, high = +80.8°C)
                       (crit = +81.8°C)
Sensor 1:     +57.9°C  (low  = -273.1°C, high = +65261.8°C)
Sensor 2:     +66.8°C  (low  = -273.1°C, high = +65261.8°C)

k10temp-pci-00c3
Adapter: PCI adapter
Tctl:         +40.5°C
Tdie:         +40.5°C

Fortunately, it never reached the 81.8C critical cutoff, but I probably need a heatsink on the new m.2 drive. Good to know when shopping next time.

On many types of workloads (except sysbench, interestingly), the new client is about 30% faster than the Wyse5070 for about 30% more cost. It also has a slightly more capable m.2 card slot, supporting both NVME and SATA cards, though it’s not clear how well they are supported by the thermals.