[Arm-netbook] A10 server dreamboard

[Gordan Bobic]

On 05/29/2012 01:24 PM, Tsvetan Usunov - OLIMEX Ltd wrote:
> Hi,
>
> I've had interesting chat with Alejandro on the irc  about possible use
> of A10 as fanless server unit.
>
> I'm interested to hear your opinion as you obviously know better this
> from customer/software developer point of view. Do you think A10 can
> handle server jobs?

Define "server jobs". I have a number of ARMs servers around, including 
the one that runs www.redsleeve.org and a bunch of other sites, mail 
server (IMAP, SMTP, SSL), mailing lists, Apache/PHP/MySQL, and all the 
other stuff that typical x86 servers do. Several smallish NAS appliances 
come with Kirkwoods in them. I am currently building a ZFS based NAS box 
based on an Armada A510 to replace a big fat Phenom X4 based one. Most 
of my server ARMs are Marvell Kirkwood (1.2GHz) based (a couple are 
newer Marvell Armada A510 (800MHz) based). A10 is certainly no worse 
than the Marvells in question.

I also use Tegra2s in my build farm and those are substantially faster 
than the Marvells mentioned.

Whether it is usable for you depends on two things:

1) Is your load such that the CPU speed doesn't matter much, e.g. a 
small web server

2) If not 1), then does your application scale horizontally? (Any 
decently architected solution should scale close enough to horizontally 
that it doesn't matter.)

If the answer to either of those is yes, then small ARM server(s) should 
be perfectly suitable.

> What do you think will be most useful (speak as performance not price)
> for ARM embedded server with A10?
> ok, ETHERNET is must, I guess HDMI and many USB ports etc are not
> necessary in this case although they are build in and will not rise the
> BOM significant

I wouldn't bother with any kind of video out for a server at all. Use a 
combined USB serial console/jtag port like on a SheevaPlug. Gate off the 
GPU and save yourself a few more mW.

> cheap LiPo 1.4Ah backup battery for stand alone operation few hours
> without main power supply instead to use costly UPS?

I wouldn't bother. If you're going to do this, use a blade type 
arrangement (or some other method of getting multiple machines in a 
single box) and have a single external UPS for the lot. Less 
complication, much more practical.

> RAM amount 1-2GB?

At least 1GB. Less than that is not worth bothering with. 2GB would be 
ideal if it can actually be made to work. I seem to remember something 
from way back then that meant that one of the bits wasn't usable which 
limits the addressable RAM to 1GB. Perhaps somebody can confirm?

> FLASH amount 4-8GB enough?

I wouldn't bother. Just use SATA.

> size factor? what is the best concept to mount in rack?

Well, my current setup uses Travla C147 (now called T1200) dual mini-ITX 
mobo chassis and two Compulab SBC-A510 motherboards. The problem with 
that setup is that it is eyewateringly expensive per node (about £800 
for the system) because the SBC-A510 boards are eyewateringly expensive.

The next one I build will probably instead cram a larger number CuBox 
boards into the same chassis (maybe with double sided foam tape, maybe 
with extra holes drilled). That should get the costs down by 3-4x for 
same spec nodes. I also tend to fit PoE splitters for power so I can run 
the whole lot from a PoE switch.

> advantage is clear, such server will have 4-5W power consumption i.e.
> less electricity bill, and quiet as no fans

Indeed, but small efficient machines already exist. Blade style setups 
also exist, but they are mind bogglingly expensive ($15K for a 16-blade 
system, approximately 10x more expensive per node than the CuBox 
solution I mentioned above).

What we really need is a blade style setup that is sensibly priced. 
$1000/node isn't anywhere near sensible. $100/node would be. If you want 
to come up with something disruptive, this is the thing to do. This is 
pretty much the only reason why I am interested in the EOMA concept - it 
would mean I could cram a whole pile of blades into a chassis that I can 
then carve up any which way required between servers, compile farms, 
etc, and still have the while thing draw < 100W.

Gordan