Low(er) power server issues
I want to put together a low(er) power soho server system. It will run Linux, act as a router, provide numerous network services and implement encrypted tunnel and disk algorithms. Mobo hardware will minimally include a 1gigabit lan and a 100Mbit wan connection, 4(+) SATA and 1 PATA. 64bit and virtualization support aren't required of the CPU.
One design goal is low power for noise (fan), heat and cost reasons. Since server idle time will be high, the system power at idle is a major factor. The server probably cannot suspend/hibernate on a regular schedule, and a 30 second "wake-up" delay is generally unacceptable for services LDAP authentication. Idling unneeded disks is a decent strategy. It's somewhat difficult to evaluate CPU idle power reqs, as well as bridge & mobo reqs, so I'm looking for experience/advise. Maybe someone knows of a website ... (silentPC hasn't been very active recently IMO).
Starting with the CPU decision ...
The current Intel Atom won't have the performance to keep up with encryption/decryption requirements at speed.
I've never been impressed wthe Celeron/Conroe series, tho' the 35W max power is attractive. Many mobile CPU parts are low power and decent performance, but the chip and mobo pricing usually make these unattractive.
The 45W Athlon X2 and LE parts should work....
Intel 65W "wolfdale" ...
AMD 65W 64 X2's ....
I'd prefer to use a passive (large) heatsink on the CPU and use a pair of case 120mm fans for system cooling and I think this should be possible for the 45/65W CPUs.
The ideal mobo would have some minimal low-end on board graphics. I intend to use the system "headless", but most mobos won't boot w/o the VGA destination.
Any thoughts ? Suggestions ?
LE is good choice but 5050e would probably work a bit better. You can easily cool them. Heat sinks without fans on them being cooled by case fans don't work too well. There's just not much mass flow of air going over them unless you can duct the case fan to within an inch or 2 of the cpu. 4850 and 5050e can be undervolted and declocked slightly to bring it's TPD down to around 20 to 25 watts.
4850e and 5050e are going to be impossible to beat for flexibility. You can declock it devolt it into your power requirement range much easier than you can overclock overvolt atom to meet the requirements. It will beat any mobile chip on processing per watt but is simply too big to use in laptops and nettops.
It'll fight voltage settings but here's pretty good estimates of power usage on a 4850e 5050e
1.6 ghz set at .9 volt wants to use about .93 volts uses about 9 to 11 watts idle and 14 15 full tilt
2.2 ghz set at 1 volt wants to use about 1.05 uses about 18 to 20 idle and 24 to 26 full tilt.
2.4 ghz set at 1.1 volts wants to use about 1.07 volts uses about 20 to 22 idle and 26 to 28 full tilt.
About double the processing power on half the power of anything LE can do because its dual core with hyperthreading. And 65nm high metal gate versus 90 to 65nm normal silicon gate.
A microatx 8200 based board or 780G based board is going to be cheapest. You could make it entirely silent by making some heatpipes out of silver or copper .375 tubing and drilling some holes in the stock cpu heatsink and 8200/780G chip and routing the tubes to some part of the case and bolting them to drive cages or motherboard trays or backside of the case or top side of the case. You'd have to remove heatsinks to remove motherboard etc but the silence may be worth it.
If you got the time and energy go crazy if you don't go easy. Just know that planning on a single 800 rpm 120mm fan and planning on no fans at all is a big jump in effort and any 120mm fan plugged into the fan cpu fan slot even with cool and quiet disabled is going to be deathly silent. My 1500 rpm vantec's that normally run about 1430 usually run about sub 1000 rpms on the cpu fan connector.
Via cpus might be good for the purpose. Many have the idle power under a watt, plus the integrated RNG, SHA and AES engines are nicely fast (plus they take no cpu away).
Nano would be a premier choice, but if not available, a recent C7 should do too. Of course it depends on if your specific encryption/decryption can be accelerated by the units (linux kernel can use the RNG, loop-aes can use the AES, there are patches for OpenSSL)
Much thanks Hephesteus - the 5050e is one I had my eye on and it should work.
Tnx too curago. My two experiences w/ Cx/Via chips have been bad. One system (Fedora 6 IIRC) could be crashed by executing certain floating point instructions. IIRC there was an errata on the cpu that the then-kernel didn't address. I doubt the CPU performance alone would be sufficient but with the crypto engine there is a good chance.
I don't mind patching OpenSSL but that addresses ssh & https web browser traffic which isn't heavy, also openvpn which appears to use SSL/TLS. Kerberos crypto for NFSv4 won't be helped. On an initial look, is seems that geode AES may be usable by ecryptfs - which would solve the encrypted FS too.
Certainly a win on the power side. I will examine this further.
[time passes ....]
I see. The AMD Geode LX series has an AES security block well supported by Linux but the CPU is slug-slow.
The Via C7-D looks like a decent fit - 1.5Ghz, and a security engine, But the Via engine support under Linux is shaky I think. Also all the C7 Mobo's for the C7 are from 3rd tier mobo makers but they are dirt-cheap (I feel the need to shower after slumming with the JetWays and BioStars). The best Mobos for C7 have 2x SATA and one IDE so they fail the cut for lack of I/O.
Most interestesting, despite the C7 1.5Ghz/20Watt point the performance seems underwhelming. If you scan http://www.smallnetbuilder.com
You'll find NAS performance comparisons for many NAS boxen. The C7 based Artigo is less than half the NAS performance of an Intel E2160 NAS. The E2160 isn't i ncontention, but according the various benchmarks the E2160 is a good 10-20% slower than the AMD 4850 that Hephasteus suggested, and slower yet that the Athlon 5050e.
If I low-volt & declock a 5050e I'd have plenty of tweak-upgradability, zero issues with compatibility/support, and clearly better performance (encryption offload aside). C7 doesn't look like a win - seems more like a lock-in to mediocre NAS performance with some open questions and patch-work about the security offload.
The idle power of a 5050e is the question I need to answer.
Last edited by stevea; 05-20-2009 at 11:08 AM.
Usefully low, I suspect: I measured the idle power of my two Linux systems at home a while back, and the AMD 5000+ was taking 60W from the wall while the dual-core Atom was taking 50W (peak obviously showed a much greater different at 109W for the 5000+ and 57W for the Atom).
Originally Posted by stevea
I was surprised by how well the AMD machine did in the comparison, and the 4850e/5050e models should have lower power consumption than that; since it spends most of the time idle but does need decent performance at times I'm probably going to replace the Atom server with a 4580e at some point and reuse the Atom for a simple desktop web-browser/email system.
Your quest is interesting to follow Please post your final results on what you end up with and how it performs
Ya have to agree with you stevea. The 4850e 5050e are pretty hard monsters to cope with by anyone wanting either silence or cheap power or what have you. I'd like to see via having a better time of things and their true random number generator stuff and cpu security stuff is pretty sweet. They are a bit behind everyone else and hopefully their Nano stuff will be a bigger winner.
But ya intel's don't like you taking their volts away but pretty much every AMD 65nm or smaller dual or quad or triple core you can take away alot of voltage and it won't even notice. I ran numbers against via's pc3 system and using a AMD 5050e won't give you quite the power sipping and the tiny package but it's a whole lot more powerful and flexible.
Geode was just a mistake and waste of time but hopefully they learned some good tech doing it.
AMD top quad comes with extra voltage stock 1.35 but the stupid thing will go from 3.2 stock to 3.7 ghz without touching voltage. They must be making these things able to work without hiccups with crazy powersupplies and voltage problems.
What's crazy with via though is the supply line. All their best stuff goes out to germany and small places in asia nobody has any of the other things. You can pick up an evergreen computer for 150 bucks but they only have their pc2 board and it's their pc3 board that is a really sweet product. And I guess the junk that has via on it they sell in america and UK are just there to make intel look good and them look bad.
If the everest had the pc3500 board in it would be fantastic buy.
But if you click on the pc2500 you will get that board in it.
Idle power depends alot on what the system has running for services. But I guarantee you using a 2.5 inch 7200 rpm laptop drive versus a 3.5" desktop drive will have more affect on idle power than anything you can cover with differences between ATOM, via, 5050e etc.
Last edited by Hephasteus; 05-21-2009 at 04:38 AM.
It will certainly help, but, if I remember correctly, the 'green' 1TB drive in my Atom server is rated at 2W when idle so there's not a lot of power you can save.
Originally Posted by Hephasteus
Just my 2 ct
Originally Posted by stevea
I dodn't have experience with headless systems yet, since I work (yet) only on x86/amd64 actively.
I don't know it arm and the others have CPUs capable of doing the jobs you want.
What I know is that the 4850e/5050e from AMD is a fine choice, it's cheap, you have a lot of Mainboards to choose from (ASUS M2A-VM ist quite nice and cheap, too) it's quite low power, passively coolable (I did exactly that), and still provides CPU power when you need it.
The other possibility I see is the mentioned VIA with padlock. The padlock part is nice the rest... huh. Well. I don't have a high opionion of it. Okay, I must admit power consumption with the right VIA ones can be really low and of course also passively coolable, but they all lack raw computing power (besides the padlock) and the drivers are often an issue. But then you didn't say you wanted graphics but since this is the main problem with drivers it might still work for you. Oh and sometimes something is broken on these mini ITX boards. In my case the HPET was in an unsuable state.
I don't know about intel but these atoms aren't that great as far as I heard.
If you go for a non x86-arch you would have to find a (meta)distribution that works on the specific arch and you won't have such a lot of packages to choose from. Gentoo e.g. has a package list comparison chart and you clearly see that most packages are x86 > amd64 > ppc and then all the others and embedded ones. On the other hand systems with power uptake less than 5W are possible.
A generally good idea would be either laptop HDD or flash memory built in. A single block of RAM (but large enough to avoid swapping or such things) and on x86 you sometimes have an undevolting option in the BIOS which can save you further watts - but check for stability.
You can run x86/amd architectures "headless"(meaning no vid display), it's the norm for servers. I've been doing so for years. A few BIOSes will allow you use use the serial port for startup/post messages, and there are a few replacement BIOSes specifically for this purpose. Otherwise the BIOS usually requires a vid card (not a display) in order to boot. Many X86 servers like the IBM X-series have a very low-end video on card on-board b/c of this. This way you only only an attached display for configuration or diagnostics.
I think a few of the higher end ARM and esp PPC CPUs might be capable; the MPC8621D for example. The problem is that I can't find any that have ~4+ SATAs + IDE + 2x enet interfaces. Usually these are built as embedded "one-off" boards for a NAS or a development card and will have at most one PCI slot in addition to some decent on-board peripherals. So for example "Buffalo Systems" might make a nice NAS system with 4 disk interfaces, but there is one enet, no PCI expansion and the amount of memory will be the minimum the vendor needed for a NAD (perhaps 64MB-128MB and soldered-in - NO mem upgrade slots. The peripherals and memory restrictions mean these are useless to me. There is also great difficult mounting these boards in a PC chassis since the form-factor is non-standard. The original chassis won't hold the peripherals I need.
I think you underestimate the Linux offerings. You can get a very complete Debian distribution for an amazing array of CPU architectures. The problem I have is getting a non-x86 system with sufficient power & mem & peripherals to do the job.
I'll post more on this later (in a rush at the moment) but I will be using some legacy drives.
The only (very few) laptop drives I have full mttf & life data on suggest these have 2x or 3x the unreliability as a 3.5inch drive. (300k-500k hours vs 750k-1500k hrs). These numbers might not be representative.
I intend to spin-down 2x of drives. This drops the power draw to around 0.8 watt each. These drives will be used for archival backup and only spun-up/mounted during periodic backup. Probably all but 1 (maybe 2) of the other drives will regularly be in "standby" which saves quite a bit even on older drives.
I agree that drive power is an issue, but there aren't many "pretty" alternatives. My intention is to grind-down my oldest drives as the "always spinning" spindles, and use a good backup/RAID policy to prevent loss when the inevitable happens.
The point of my effort is to save money, not to save the hypothetically endangered polar bears at a premium. I estimate I'll pay ~$1/watt-year unless the current regime imposes cap&trade on carbon, in which case the figure might become $1.50/watt-year. If I can save ~25-30 watts continuously that's worth about $125-$150 over a 5 year server lifespan. It's NOT worthwhile for me so spend say $400 dollars in new disks for the privilege of saving $50 on disk power. I already own a lot of fully amortized disks. It's about frugality. ... but that calculation needs to be made.
Last edited by stevea; 05-24-2009 at 10:42 PM.