Archives

Categories

HP ML110 Gen9

I’ve just bought a HP ML110 Gen9 as a personal workstation, here are my notes about it and documentation on running Debian on it.

Why a Server?

I bought this is because the ML350p Gen8 turned out to be too noisy for my taste [1]. I’ve just been editing my page about Memtest86+ RAM speeds [2], over the course of 10 years (high end laptop in 2001 to low end server in 2011) RAM speed increased by a factor of 100. RAM speed has been increasing at a lower rate than CPU speed and is becoming an increasing bottleneck on system performance. So while I could get a faster white-box system the cost of a second-hand server isn’t that great and I’m getting a system that’s 100* faster than what was adequate for most tasks in 2001.

HP makes some nice workstation class machines with ECC RAM (think server without remote management, hot-swap disks, or redundant PSU but with sound hardware). But they are significantly more expensive on the second hand market than servers.

This server cost me $650 and came with 2*480G “DC” grade SSDs (Intel but with HPE stickers). I hope that more than half of the purchase price will be recovered from selling the SSDs (I will use NVMe). Also 64G of non-ECC RAM costs $370 from my local store. As I want lots of RAM for testing software on VMs it will probably turn out that the server cost me less than the cost of new RAM once I’ve sold the SSDs!

Monitoring

wget -O /usr/local/hpePublicKey2048_key1.pub https://downloads.linux.hpe.com/SDR/hpePublicKey2048_key1.pub
echo "# HP monitoring" >> /etc/apt/sources.list
echo "deb [signed-by=/usr/local/hpePublicKey2048_key1.pub] http://downloads.linux.hpe.com/SDR/downloads/MCP/Debian/ stretch/current-gen9 non-free" >> /etc/apt/sources.list

The above commands will make the management utilities installable on Debian/Buster. If using Bullseye (Testing at the moment) then you need to have Buster repositories in APT for dependencies, HP doesn’t seem to have packaged all their utilities for Buster.

wget -r -np -A Contents-amd64.bz2 http://downloads.linux.hpe.com/SDR/repo/mcp/debian/dists

To find out which repositories had the programs I need I ran the above recursive wget and then uncompressed them for grep -R (as an aside it would be nice if bzgrep supported -R). I installed the hp-health package which has hpasmcli for viewing and setting many configuration options and hplog for viewing event log data and thermal data (among a few other things). I’ve added a new monitor to etbemon hp-temp.monitor to monitor HP server temperatures, I haven’t made a configuration option to change the thresholds for what is considered “normal” because I don’t expect server class systems to be routinely running above the warning temperature. For the linux-temp.monitor script I added a command-line option for the percentage of the “high” temperature that is an error condition as well as an option for the number of CPU cores that need to be over-temperature, having one core permanently over the “high” temperature due to a web browser seems standard for white-box workstations nowadays.

The hp-health package depends on “libc6-i686 | lib32gcc1” even though none of the programs it contains use lib32gcc1. Depending on lib32gcc1 instead of “lib32gcc1 | lib32gcc-s1” means that installing hp-health requires removing mesa-opencl-icd which probably means that BOINC can’t use the GPU among other things. I solved this by editing /var/lib/dpkg/status and changing the package dependencies to what I desired. Note that this is not something for a novice to do, make a backup and make sure you know what you are doing!

Issues

The “HPE Dynamic Smart Array B140i” is a software RAID device. While it’s convenient for some users that software RAID gets supported in the UEFI boot process, generally software RAID is a bad idea. Also my system has hot-swap drive caddies but the controller doesn’t support hot-swap. So the first thing to do was to configure the array controller to run in AHCI mode and give up on using hot-swap drive caddies for hot-swap. I tested all the documented ways of scanning for new devices and nothing other than a reboot made the kernel recognise a new SATA disk.

According to specs provided by Dell and HP the ML110 Gen9 makes less noise than the PowerEdge T320, according to my own observations the reverse is the case. I don’t know if this is because of Dell being more conservative in their specs than HP or because of how dBA is measured vs my own personal annoyance thresholds for sounds. As the system makes more noise than I’m comfortable with I plan to build a rubber enclosure for the rear of the system to reduce noise, that will be the subject of another post. For Australian readers Bunnings has some good deals on rubber floor mats that can be used to reduce server noise.

The server doesn’t have sound hardware, while one could argue that servers don’t need sound there are some server uses for sound hardware such as using line input as a source of entropy. Also for a manufacturer it might be a benefit to use the same motherboard for workstations and servers. Fortunately a friend gave me a nice set of Logitech USB speakers a few years ago that I hadn’t previously had a cause to use, so that will solve the problem for me (I don’t need line-in on a workstation).

UEFI and Memtest

I decided to try UEFI boot for something new (in the past I’d only used UEFI boot for a server that only had large disks). In the past I’ve booted all my own systems with BIOS boot because I’m familiar with it and they all have SSDs for booting which are less than 2TB in size (until recently 2TB SSDs weren’t affordable for my personal use). The Debian UEFI wiki page is worth reading [3]. The Debian Wiki page about ProLiant servers [4] is worth reading too.

Memtest86+ doesn’t support EFI booting (just goes to a black screen) even though Debian/Buster puts in a GRUB entry for it (Debian bug #695246 was filed for this in 2012). Also on my ML110 Memtest86+ doesn’t report the RAM speed (a known issue on Memtest86+). Comments on the net say that Memtest86+ hasn’t been maintained for a long time and Memtest86 (the non-free version) has been updated more recently. So far I haven’t seen a system with ECC RAM have a memory problem that could be detected by Memtest86+, the memory problems I’ve seen on ECC systems have been things that prevent booting (RAM not being recognised correctly), that are detected by the BIOS as ECC errors before booting, or that are reported by the kernel as ECC errors at run time (happened years ago and I can’t remember the details).

Overall I’m not a fan of EFI with the way it currently works in Debian. It seems to add some of the GRUB functionality into the BIOS and then use that to load GRUB. It seems that EFI can do everything you need and it would be better to just have a single boot loader not two of them chained.

Power Supply

There are a range of PSUs for the ML110, the one I have has the smallest available PSU (350W) and doesn’t have a PCIe power cable (the one used for video cards). Here is the HP document which shows the cabling for the various ML110 Gen8 PSUs [5], I have the 350W PSU. One thing I’ve considered is whether I could make an adaptor from the drive bay power to the PCIe connector. A quick web search indicates that 4 SAS disks when active can take up to 75W more power than a system with no disks. If that’s the case then the 2 spare drive bay connectors which can each handle 4 disks should be able to supply 150W. As a 6 pin PCIe power cable (GPU power cable) is rated at 75W that should be fine in theory (here’s a page with the pinouts for PCIe power connectors [6]). My video card is a Radeon R7 260X which apparently takes about 113W all up so should be taking less than 75W from the PCIe power cable.

All I really want is YouTube, Netflix, and text editing at 4K resolution. So I don’t need much in terms of 3D power. KDE uses some of the advanced features of modern video cards, but it doesn’t compare to 3D gaming. According to the Wikipedia page for Radeon RX 500 series [7] the RX560 supports DisplayPort 1.4 and HDMI 2.0 (both of which do 4K@60Hz) and has a TDP of 75W. So a RX560 video card seems like a good option that will work in any system that doesn’t have a spare PCIe power cable. I’ve just ordered one of those for $246 so hopefully that will arrive in a week or so.

PCI Fan

The ML110 Gen9 has an “optional” PCIe “fan and baffle” to cool PCIe cards (part number 784580-B21). Extra cooling of PCIe cards is a good thing, but $400 list price (and about $50 ebay price) for the fan and baffle is unpleasant. When I boot the system with a PCIe dual-ethernet card and two PCIe NVMe cards it gives a BIOS warning on boot, when I add a video card it refuses to boot without the extra fan. It’s nice that the system makes sure it doesn’t get into a thermal overload situation, but it would be nicer if they just shipped all necessary fans with it instead of trying to get more money out of customers. I just bought a PCI fan and baffle kit for $60.

Conclusion

In spite of the unexpected expense of a new video card and PCI fan the overall cost of this system is still low, particularly when considering that I’ll find another use for the video card which needs and extra power connector.

It is disappointing that HP didn’t supply a more capable PSU and fit all the fans to all models, the expectation of a server is that you can just do server stuff not have to buy extra bits before you can do server stuff. If you want to install Tesla GPUs or something then it’s expected that you might need to do something unusual with a server, but the basic stuff should just work. A single processor tower server should be designed to function as a deskside workstation and be able to handle an average video card.

Generally it’s a nice computer, I look forward to getting the next deliveries of parts so I can make it work properly.

Comments are closed.