[tahoe-dev] expanded testnet

[Brian Warner]

Thanks to Zandr, we've just added four new server boxes to our public
testnet. These machines are all professionally-hosted colocated machines,
with big pipes and large disks. One machine is being used (underutilized,
really) as the Introducer/vdrive-server, and the other three are being used
for storage servers. We still have one of the old machines as a storage
server as well.

The new machines have four disks apiece, and we're running one node per
spindle. This gives us a total of 13 in-colo storage server nodes, so our
3-of-10 encoding is finally able to place each share on a different disk.
I've moved three of the four nodes that used to be on the old machine onto
distinct new ones, leaving one in place, so now the files that were uploaded
before yesterday should be spread across the new machines (although not
uniformly across the new disks.. reliability for those files will not be
quite as high as for new ones).

The new machines are also noticeably faster than the old ones, enough to make
our automatically-measured upload/download rates go up. The old machines were
giving us 1.58MBps up / 4.09MBps down, while the new ones get us 1.87MBps up
/ 4.42MBps down. Check out the step function yesterday (17-Oct-2007) at 16:30
on our munin performance graphs:

http://allmydata.org/tahoe-figleaf-graph/hanford.allmydata.com-tahoe_speedstats_rate.html

Those numbers will probably improve some more when I distribute the perf-net
(the separate network we use just for performance testing) over all the new
machines. At the moment it is running entirely on a single box, so we're
probably seeing some contention between the different storage servers. I
think we're CPU bound on those tests right now, rather than disk or IO bound
(disks should be able to sustain transfer rates larger than what we're
seeing, and these machines are connected by either 100Mb or 1Gb network
links).

cheers,
 -Brian