Got a power hungry server

[prppedro]

Hey, people

It's a silly question, I admit.

I got an old HP ML350 G5 sporting two Xeons 5120. After formating it to Linux and removing both the PPM and the second processor, Kill-a-Watt shows 178~180W, for the high load generated by Prime95, and 158~162W while sitting idle. Well, according to this site¹, 5120 burns between 6W (best case, recent revision) and 12W (worst case, old revision) while idling away. iLO 2 have been switched off as well. So, processor seems not to be the problem. The two old SATA drivers burns a good ammount as well, but never more than 20W~30W.

So, the rest of the hardware are burning a lot of power. What gives?

Is there any way of cutting it down a little bit? I've been looking for L or LV Xeons with lower TDP's, but I'm afraid it wouldn't help much either.

[1] https://techreport.com/news/13036/new-x ... dle-power/

Thanks!

[CwF]

Interesting subject! Another reason I prefer Intel!
I'm not so sure your cpu wattage is accurate, but your right that they are never the whole story. I've done lots of hardware testing and building and usually are looking at thermals rather than wattage, but a similar thing really. Overlooked things are many. The power supply can be a variance, up to 50w+. Extra cards and drives obviously. The chipset and memory also vary. Within a generation the chipset choices can range 30-50w or more, and memory config can make a big difference.

Don't know on your system, but an oversized heavy duty power supply could be +10-20% of what it 'could' be. Older 5v memory, especially if many lower capacity dimms could be much of the draw.

On a server chipset, there may well be many options in the bios. Usually they would be set to stay awake, so power save, idle states C0 and the like might be turned off. You may have control over the base clock, most xeons do, so you can lower it without effecting turbo speed. Usually the more you set for low idle the more sluggish the machine becomes.

[kedaha]

It may not be too difficult to come by an old, second-hand server like that; folks probably get shot of them precisely because they are rather power hungry, specially when they're on round the clock day in, day out. An old server like that might be profitable for certain production purposes but if it's just for use, for example, as a home web & email server, then I'd get rid of it and use something which hardly used any power at all, like an sbc, which is the subject of my latest topic.

[No_windows]

I don't recall what mine is pulling, but I think it's similar numbers.... it's a DL380 G5 with two processors. Don't forget you've got a lot of fans in there for cooling all those components in such a compact space. The raid card gets VERY hot, so that's not helping either.

I pulled one of the power supplies, that was the extent of my power savings efforts.

I've been looking at local auctions for a replacement, but I haven't had the money when something decent has come up. I may need to just get a PC and be done with it.

[CwF]

I favor the server end, but much more modern! I like the 3.5SBC's on the more powerful side of things for mobile use, but for home I try my best to imagine all the little gadgets I could build - then re-imagine them into a bunch of vm's - then put them all in one big wicked box. Once there is always something going on, one box is the most efficient.

[pylkko]

Couldn't you just use containers, then?

[CwF]

Me, Containers? No.
Since containers don't provide everything I need, they add nothing but another layer of complexity.
If one can get away with only the container layer, then not need the qemu complication layer, then by all means, containerize! I see no benefit from using both if one will do, and vm's do more, so that's it.

[prppedro]

Well, thanks, guys, for all the replies.

To be quite honest, I really liked the case (Must admit, I find all these enterprise stuff really gorgeous, and they're rigged as hell). Problem is, it's all proprietary stuff from PSU connections down to motherboard form factor. Therefore, it's a no go fitting a common consumer grade µATX inside it. So, given it's my first server and I don't want to just throw it right away, I might well turn it in a backup server, so it lights on only when I need it to. The much more sane approach presented by kedaha is probably what I'll try to do for the 24/7 service.

But it's a bit intriguing to me how much power it spends compared to another machine I have sporting a Xeon X5460 jury rigged to a LGA 775 board, the P5QC. It's a terrific high end (albeit consumer grade) board. Powered by an Aerocool 80Plus White PSU, it sit's at 80W while idling. While blasted by Prime95, it hovers between 150W~165W, which my ProLiant only achieves while idling.

So, It's probably something to do with high end stuff included in the board, like FBDIMM memory controllers, redundant circuitry and all that jazz... Or even the PSU, as some of you pointed out, as well. But I'm actually surprised it makes that difference.

P.S.: CwF suggests using virtualization. Well, that's one of the things I hoped to run. I run several things on my E3-1245 v2, on my main rig. None of them on a permanent basis, because they tend to clog up the machine resources, which makes browsing and reading a little bit sluggish, at times. So, for things that are kind of start and forget, I'd like to run on a separate server.

[CwF]

@prppedro
I feel the need to qualify the scale of things. and scale makes all the difference.
Power consumption is not linear to performance. Computers have made more gains in power usage than outright performance. Cores are part of this. So a computer 10X your E3 is possible, and it doesn't draw 10X power. I have an E3 that has a full range from a 46W idle to about 90W. It's performance is matched on an E5 that will idle around 65W and peak over 200. The E3 emulation on the E5 (doing the same task) is about a 35W premium over what else is going on. So, boot the vm and add a minimal overhead, or boot a separate machine that will suck 50+ extra watts.

I've said elsewhere, up the ante. Once the machine can run a handful of full power E3 scaled VM's, something minor like a file server as a vm is literally a few watts plus the extra disk, a fraction of a dedicated nas box. Once there are 4-5 'little' things going on the big box methods wins. Once there are 10+, it's not even fair.

[prppedro]

@prppedro
I feel the need to qualify the scale of things. and scale makes all the difference.
Power consumption is not linear to performance. Computers have made more gains in power usage than outright performance. Cores are part of this. So a computer 10X your E3 is possible, and it doesn't draw 10X power. I have an E3 that has a full range from a 46W idle to about 90W. It's performance is matched on an E5 that will idle around 65W and peak over 200. The E3 emulation on the E5 (doing the same task) is about a 35W premium over what else is going on. So, boot the vm and add a minimal overhead, or boot a separate machine that will suck 50+ extra watts.

I've said elsewhere, up the ante. Once the machine can run a handful of full power E3 scaled VM's, something minor like a file server as a vm is literally a few watts plus the extra disk, a fraction of a dedicated nas box. Once there are 4-5 'little' things going on the big box methods wins. Once there are 10+, it's not even fair.

Ehhh... I don't know. Indeed, there are E5 chips that are pretty much power hogs. And I know for sure TDP doesn't translate into computing power (X3430 is a 95W processor and loses to L3436 in PassMark). My problem with using my E3 to do this kind of thing is that it's my desktop machine. It's a rare event, but something things go south, X crashes the entire system, memory gets full... So I don't like much running everything in the same hardware as I use to browse, develop, study etc.

On other hand, those services aren't exactly heavy things. I need a backup system, SSH, DCHP/DNS, database and firewall. They're a bit too heavy for Pi, tho. Probably the best solution would be buying an Atom or Celeron board anyways.

[pylkko]

Me, Containers? No.
Since containers don't provide everything I need, they add nothing but another layer of complexity.
If one can get away with only the container layer, then not need the qemu complication layer, then by all means, containerize! I see no benefit from using both if one will do, and vm's do more, so that's it.

Yeah, I meant containerized services as opposed to having individual VM images (entire OS and all) for each service.

. I need a backup system, SSH, DCHP/DNS, database and firewall. They're a bit too heavy for Pi, tho. Probably the best solution would be buying an Atom or Celeron board anyways.

So, how does my MIPS-based router do all that then? It's way more pathetic than a pi....

[kedaha]

So, given it's my first server and I don't want to just throw it right away, I might well turn it in a backup server, so it lights on only when I need it to. The much more sane approach presented by kedaha is probably what I'll try to do for the 24/7 service.

Just to say, I've recently hooked up my old Acer Aspire One, which only has 1G RAM to to the router in order to try a YunoHost server. I've written about this in another forum topic about running a home server. Energy consumption, of course, is negligible and the nginx server works lightning fast. Now I've notified my dedicated server hosting provider that I'm terminating the contract when it expires; it's not quite the same but one can run a multi-domain personal server at home with the full panoply at minimal cost.