The current kernel 3.5 from rabeeh and vDorst miss the patch from 220.127.116.11
that turn off hardware checksum when the mtu is larger than 1600. When you
set mtu to 9000 using ifconfig or set it in /etc/network/interface, the throughput
will increase dramatically. With normal file transfer no encryption, I can get 70-90 MB/s.
For "benchmark" number, iperf can go over 100MB/s easily as a client or server with
the jumbo frame patch. Without jumbo frame patch, iperf can go around 77MB/s.
See the jumbo frame patch from wikipage:http://www.solid-run.com/mw/index.php/L ... ic_patches
I have successfully run glusterfs on two cuboxes. Each box has this config:
* Latest kernel 3.5 from vDorst with jumbo frame patch
* Debian Wheezy armhf
* glusterfs 3.3 (compiled with gcc 4.7)
* 2 hard drives, one 1TB, one 2TB in Mediasonic Probox 4bay, connected to CuBox by esata cable.
What I observe is the throughput of each CuBox is around 28-35MB/s with 77-89 cpu utilization
from top command.
My client runs on an Intel IvyBridge box with Debian Wheezy.
Perhaps some fine-tuning in tcp buffer to avoid buffer bloat?
A faster CuBox, some sort of quad-core A9 1.2-1.5GHz? A fast enough cpu with SATA or eSata
AND Gigabit Ethernet will be an awesome storage solution. Think of low-power "cloud storage"
at home or for small business.