Adventures chasing a quiet rack
I’ve seen this same query brought up a few times on various forums, so it seems I’m not the only person with this query. Like many nerdy types, I’ve wanted a home datacenter as long as I’ve known about datacenters. In the early 2000s, I did the “ATX towers on a breadrack” thing, and when I started looking for a machine to do home server duties here I didn’t think too hard and ended up splurging on a PowerEdge R510.
At first I thought the noise was fine - on the first boot on my kitchen table it sounded like (as I’ve described many times before) “9 hairdryers in a toolbox all turned on at once” and blew the styrofoam packing off the table and across the kitchen, but it quickly settled in in all but the summer months and I was okay with having it in my office. At least, I thought I was.
Fast forward two years, and it begins driving me batty. I tried putting it in a closet - this worked for a few hours, until the lack of ventilation drove temps up a bit enough to cause the fans to start ramping, and this combined with it hard up against a solid wooden shelf that reverberated meant the result was actually louder. I found an enclosed (but mesh-front) rack for cheap, and set about trying to shut the fucking thing up.
Oh, and it’s worth noting that this was not long after I spent a bundle of money on my desktop making it near-silent during gaming.
The software solution
There are numerous threads and articles on using IPMI to manually control the fans, ramping them down below the minimum speed that iDRAC would tend to select. I didn’t like this idea, mainly because my office at the time got pretty damn hot in the summer, but it does definitely do what it says on the tin.
Hardware solution #1: Soundproofing the rack
I bought some of the cheap, “foam pyramid” “soundproofing” treatment on Amazon, and two cans of Sikaflex adhesive, and set about taking my rack apart. I carefully trimmed the entire thing, including the mesh front and back, and reassembled everything. The results were… disappointing. It certainly changed the tone, taking a tiny bit of the whistle off the top end, but the actual difference in sound levels were below the noise floor of what an advertising-supported Android application on a Motorola Moto G4 could measure. I’m honestly not convinced it was any quieter at all, once the benefit of confirmation bias (my brain protecting me from realizing I just spent about $60 on foam that didn’t do a fucking thing) wore off.
Thinking back to my days in car audio, where I had none of the gear and no idea, but pretended I did and spent a lot of effort doing things on a budget to some effect, I decided that what I really needed was some mass and some direction changes if I wanted to mute the sound. So I left the sound deadening foam on the larger panels on either side, and on the mesh front and back I tore it off, replacing it with half-inch thick MDF cut to the exact shape.
A small piece of board (later replaced with acrylic, though I should just spend the money on blanking panels) to separate a cold and hot side, and a pair of 120mm fans in the top where the cable bundles are supposed to pass in and it did actually make a difference. It was still definitely audible, but I could stand to be in the same room as it for a time, and surprisingly there was zero difference in temperatures.
As it was picking up air from the bottom of the rack in order to blow through the server though, there was an increase in dust… the rack sitting on carpet and all. This didn’t bother me, because dusting the server as part of a service is already a twice-yearly routine for me anyway.
But I could still hear the bloody thing, and it was starting to annoy me.
Hardware solution #2: Replace the R510
I found what seemed to be a pretty good deal on an old Supermicro chassis - specifically an SC835, not to be confused with the much more desirable SC836, but still a good unit for it’s age. I was sacrificing two or four (depending how you counted them) hotswap bays, but considering I was only using 3 at the time I could live with that.
It sounded good to me when I picked it up, so I bought it. When I got it home though and compared them side by side, initially it was no quieter on boot, and it actually didn’t ramp the fans down as quiet as the Dell did, so I was pretty upset.
However I quickly (for some definition of the word, some parts took months) realized a few factors:
- The power supply is hideously noisy, and there are two. Unplugging one works, as long as it’s not pushed all the way home because the fault alarm will sound if it’s attached and not powered. Better still, replacing the one I was using with the “SuperQuiet” model helped a lot.
- The fans themselves don’t ramp down low enough, for my environment, even in the summer, they spin way faster than they need to. I tried various methods to fix this in software to no avail, and ended up taking the advice of a random comment on GitHub and connected each one via “low noise adaptors” (basically an inline resistor on the supply pin). The resistors got hot, but it did work.
- The processors that came with this machine were older and less efficient than the ones that were in the R510. As I was unable to sell the R510 for a price that I was happy with, I swapped the CPUs out, and this improved things further - CPU core temps dropped about 5~10C and I picked up more processing power to boot (not that I needed it, my needs are very modest). I considered “low power” CPUs, but they don’t appear to change the idle power usage, they just put a ceiling on the off-idle TDP, and thus for my mostly-idle application would be a waste of money.
- Eventually, I realized (after burning myself on it) the temperature indicated on the chipset was not a misreading, and I put an old 80mm fan on it and immediately the fans ramped down further and stayed there.
Hardware solution 2.5: Replace the Xeons
The main thing that the Supermicro chassis did gain me was the ability to use a commodity ATX board. I ummed and ahhed over what to use, almost pulling the trigger on an Asrock Rack Ryzen board before wimping out. This would ultimately have been the best way forward, as all I’d need to do is ensure that the fans have sufficient airflow and static pressure to keep the hard drives cool, and everything else would be so much lower TDP that the fans moving air over the disks would cover it.
I didn’t end up going this way as it looked like we were about to buy a house, so I tightened my purse strings and put up with the noise.
The real solution: get this thing the fuck away from me
After we moved into the house we bought in 2021, I finally had an attached, finished garage that I was able to move the networking equipment and servers into.
I removed the wood panelling, leaving the front and rear mesh completely open, and I removed all the “low noise adaptors” from the fans. The server’s noisy, but who gives a shit because no one is near it to hear it for very long.
Dust is now a concern, but again I do regular maintenance including blowing all the dust out with a Datavac so I think I can live with that.
We’ve not yet been through a summer, but judging from the winter temps I think it will be controllable. Modern server gear doesn’t need to be run as cold as datacenters in the early 2000s were, but it’s a hard thing to let go of for sure… the idea that this machine’s going to overheat and cook. The primary concern is the disks, as always - but I have temperature monitoring in place to keep an eye on these things.
But the main factor is that I am optimistic the garage will prove to have fairly reasonable temperatures. It’s a good sized space, and completely finished - insulation and everything (though the insulation is not great and missing in places, but it won’t be super expensive to fix that). I think with some ventilation to move the worst of the air around and it’ll be fine for most of the year.
But as far as noise levels in my office go, words can’t express how nice it is that my MacBook is for the most part the noisiest thing in here. I thought I was okay with the whirr of fans, because that’s what nerds do, but in retrospect that is not the case.