[svlug] Sheevaplug ?

Rick Moen rick at linuxmafia.com
Wed Oct 21 14:16:33 PDT 2009


Quoting Benjamin Floering (floering at ieee.org):

> Yes, I have it and love it.  I received mine back in April and have since
> migrated all of my persistent operations in my datacenter to the plug.  It
> came pre-installed with Ubuntu and worked great out of the box, but I
> removed that and put gentoo on it and the latest kernel with support for
> lower power operation.  I have a 1TB USB hard drive attached to it and it
> works wonderfully as a NAS for video and music streaming to various devices
> around the house.  It also functions as a backup device, bittorrent
> downloader, DNS cache, VPN server, and VPN client constantly connected to my
> high bandwidth co-loc basically extending my network to include the
> production webserver. 

I am interested, but have a concern.  Maybe you would be good enough to
comment.

The Sheevaplug uses a Marvell 88F6281 SoC (system on a chip), which is 
their variant on an ARMv5TE-family design, reportedly using some of what
they got from acquiring Intel's XScale division.

Marvell's data sheet is here:
http://www.marvell.com/files/products/embedded_processors/kirkwood/HW_88F6281_OpenSource.pdf

Looking through the hardware specs for general-purpose FPU (floating
point unit; math chip) functions, I find:

Security Engine:  AES, DES, 3DES, SHA-1, MD5
XOR Engine, qty 2
MPEG2-TS and some audio support (Sony-Philips Digital Interconnect
   Format audio, and Integrated Interchip Sound.


There's nothing in the specs about general-purpose floating point.
That's where I wonder about the system bogging down if used for
(especially) multiple SSL and SSH functions, plus gpg, and so on.

In theory, crypto libs could be compiled to access the particular
hardware-supported ciphers via their hardware interfaces, _given_ the
assumption of those ciphers being of primary importance.  I certainly
like AES and 3DES, but RSA and DSA (to name two) are very prevalent and
also very computation-intensive, not to mention (on the symmetric-cipher
side) Blowfish, Twofish, etc.  Also, MD5 hashing's going away, and
SHA-1's becoming a bit dicey.  (There will be a successor, to be called
SHA-3, in a couple of years.[1])

The point being that hardware-supported specific ciphers are nice if 
common crypto libs talk to them, _but_ are not going to cover general 
crypto usage -- let alone that of the near future.

The last time I made significant use of a machine lacking
general-purpose FPU on Linux was a friend's Corel Netwinder -- which
coincidentally also used an ARM-family CPU (Digital Semiconductor
SA-110).  Of course, I'm not saying this was the same thing (being circa
1998, if nothing else), but I remember that X11 drawing was sometimes
painfully slow, depending on what object the X server was rendering, of
course.  It was really obvious, anyway, that the lack of FPU (and
consequent need to emulate an FPU entirely in software) had a
serious performance cost for anything computation-intensive.

So, if you can comment on how _your_ Sheevaplug handles varied crypto
under significant system load, that would help the rest of us assess 
that option.  Thanks.

I would _love_ to retire my PIII-based server in favour of something
that draws 5 Watts, plus a pair of USB-connected external 2.5"
(mirrored) 1 TB SATA hard drives drawing whatever they draw, but not if
it becomes dog-slow under mixed SSH/SSL/gpg loading.

I note that some of the newer AMD Athlon II CPUs are claimed to
draw 45 Watts while doing "full multimedia performance", which is better
than a kick in the pants if true, and you're at least guaranteed to 
have full FPU.  (That 45 Watts is not the motherboard or complete
system, just the CPU.)  

I had high hopes for new 1U systems using Intel Atom hitting ARM-type
power-thriftiness levels, but Luke Crawford tried one of those and found
that the motherboard chipset draws only 20% less power than a comparable
system using Intel Core 2 Duo.
http://linuxmafia.com/pipermail/conspire/2009-October/005036.html

[1] http://en.wikipedia.org/wiki/NIST_hash_function_competition






More information about the svlug mailing list