13W idle Atom N270 server for 236 euros

tuomaspt · Post by **tuomaspt** » Mon Jul 05, 2010 6:13 am

An old Athlon XP based server started to have troubles booting up, and a replacement of condensers didn't help, so it was time to pick up the wallet and go shopping.

Components and price paid (including possible shipping):

Intel D945GSEJT Atom N270 motherboard 112.70e
Western Digital Scorpio Blue 320GB 2.5" 50.90e
MW3H36GS 36W switching power supply 24.68e
Kingston KVR533D2S4/2G 2GB 533Mhz CL4 SO-DIMM 48.00e
---
Sums up to about 236 euros.

Beside the lowest powered Intel Atom I considered a SheevaPlug computer or similar, but they come with only 512MB of system memory, which was too little for me. Also the shipping costs would have made it more expensive than the Intel Atom board. The benefit would have been far lower power usage.

The system was built in an old ATX case I had laying around. The PSU sized hole in the following picture is just big enough to fit a hand in to check how hot the heatsinks get (not too hot to keep a finger on when idle):

Next the motherboard. This box is firewall, amongst being a web, mail, media and many other things server, and so it needs two NICs. The extra NIC fills up all the PCI expandability, except for a mini PCIe slot. After I upgraded the BIOS to the latest one the board started to use the memory at CL4 as specified instead of CL5.

The hard-drive fastened with 2.5" -> 3.5" brackets I had. This is the only component that makes noise in the system. First I tried to fit in an old 3.5" ATA drive, but despite the manual saying it should work, I wasn't able to get the motherboard to recognize it. The drive was unresponsive and made awful head parking clicking sounds many times in a minute until I disabled the faulty "advanced power management" features with "hdparm -B 255". This didn't raise the measured idle power consumption.

The power supply. Specified MEPS: IV and CEC: IV and EUP. After much unnecessary waste of time, I found this means it has specified minimum efficiency of about 82-83%. I believe this is true at loads at least down to 20% of total DC output, but the specifications are not that clear. The PSU makes annoying whine when nothing is plugged to it and it is connected to the wall. Otherwise it is silent.

Power usage:

- (13 +- 1)W draw from wall socket at idle measured with an energy meter.
- (19 +- 1)W maximum peak load observed so far.

Noise:

- The HD makes some noise.

Temperatures:

- Temperature inside the case barely rises from ambient (measured).

Software migration:

I was able to use the Gentoo Linux installation on the old Athlon XP system directly, because both systems are 32-bit i686. The N270 CPU does not support 64-bits. I copied the data from the old harddrive to the new harddrive, configured a new kernel and it was ready.

I like (+):

- Relatively low power consumption both idle and load
- Does not heat up a closet
- Noticeably faster than the Athlon XP (900Mhz) system it replaced
- Easy software migration

I dislike (-):

- Irregular SO-DIMM memory and incompatibility of the old HD increased costs
- Pretty much non-expandable unlike standard ATX PC
- The northbridge and southbridge have high power draw compared to the CPU.
- Temptation to buy a neat small case for it

mark19891989 · Post by **mark19891989** » Sun Jul 11, 2010 1:04 pm

nice build, how a personal server should be

in a case this size, i would be so tempted to pack it full of 2tb drives, but dont think you have enough sata ports .

you said you store in the cupboard, so there isnt really much point in getting a nice looking case then hide it out of sight.

what services are you running on this build?

i have also been tempted to get a SheevaPlug , but in the end i just found doing a regualr build gave me more flexability.

for that price, 13w idle is nice, my server is at about 30w idle, 75w load, so your consumption is alot better

ntavlas · Post by **ntavlas** » Sun Jul 11, 2010 10:52 pm

Nice and neat. I`m also working on a few builds using this board and it`s amazing how little heat it produces along with a notebook hard drive.

Is it just a simple samba file server or a gateway too (looking at the second net cable)? I`m tempted to use one of those as a server, what kind of transfer speeds do you think I could expect?

Thanks for sharing

tuomaspt · Post by **tuomaspt** » Mon Jul 12, 2010 10:01 am

It runs a firewall, Apache, Zope, MySQL, NFS NAS and several other services.

Copying a 350MB file from Apache I got 55,3MB/s (measured once, according to wget), the bottleneck being possibly the harddrive. However, Apache did seems to take nearly 100% process time in top during the operation. Copying a 3GB file over NFS I got 36MB/s (speed of the first try). With NFS I noticed during the copying the server spawned many threads/processes, which seemed to take all the CPU according to top. Also on the destination machine network activity measured roughly 40-60MB/s during the copying so there was quite some inefficiency involved. With SFTP I measured transfer speed of roughly (12.2 +- 0.2) MB/s (according to the figures SFTP gives). I hope that helps.

ntavlas · Post by **ntavlas** » Mon Jul 12, 2010 6:51 pm

That helps, thanks. I`m planning to run samba though I hoped for something better than 36mb/s. Still this seems to be the norm, another member running samba on the same motherboard got very similar speeds. On the other hand I've seen faster systems reach close to 80mb/s which makes me think that the cpu is indeed the bottleneck. As it is, it seems impossible to build a fast nas that is so power efficient, unless someone turns a laptop into one.

tuomaspt · Post by **tuomaspt** » Tue Jul 13, 2010 12:25 am

Noting the inefficiency I redid some NFS benchmarks now with async directory export setting instead of sync, which is known to greatly affect performance.

I run the bechmark by issuing "time cp /mnt/nfs_dir/file /dev/null" on the client machine. I divide the size in bytes (obtained from 'ls -l') by time in seconds printed by 'time'. These are files that are unlikely to be in any cache.

1,5GB file 48,4MB/s
1,3GB file 50,6MB/s
1,2GB file 48,6MB/s
0,4GB file 63,4MB/s
0,7GB file 59,6MB/s

I think it is, however, impossible to measure the throughput of the network like this as I have seen anything between 40..70MB/s as throughput of this harddrive in some benchmarks on the web.

batka · Post by **batka** » Mon Jul 19, 2010 7:45 am

Good setup!
Maybe worth to replace the casing!

Is your config maybe enough for 8x DVD burning? DVD data rate at 8x speed is 11Mbyte/sec theoretically. So I guess there should be working nicely.....

lifeisgood · Post by **lifeisgood** » Mon Jul 19, 2010 8:56 pm

tuomaspt · Post by **tuomaspt** » Sat Jul 24, 2010 3:48 am

I don't see any reason why one couldn't burn a DVD with it as long as the harddrive can feed it data fast enough. The WD notebook drive can certainly sustain the 11MB/s needed for 8x. I can't test it for you, however

Gentoo linux. When GCC 4.5 goes stable on Gentoo I will do a full system recompile optimized for the atom CPU.

tuomaspt · Post by **tuomaspt** » Wed Jul 28, 2010 9:41 am

As adviced in another thread I ran "iperf" to measure network throughput regardless of harddrive speed.

I am using the onboard NIC and the connection goes through a D-link 1GB network switch to onboard NIC on Asus P5Q. On the server I run "iperf -s" and on the client first "iperf -c serverip" four times and then I specify different TCP window sizes with for example "-w64k" a couple of times. The server is not completely idle, but not really running any cpu intensive process either during the benchmarks.

TCP window size / Megabytes per second

16k 88.4MB/s
25.3k 89.8MB/s (for some reason it gave 25.3k this time)
16k 89.8MB/s
16k 89.5MB/s
128k 89.5MB/s
256k 89.8MB/s
64k 87.5MB/s
64k 90.3MB/s