[SLL] Demonstration next month (was: ATA-over-Ethernet 0x88a2)

[Russell Evans]

On Fri, 6 May 2005 19:27:19 -0700 (PDT)
"Andrew Sweger" <andrew at sweger.net> wrote:

> On Fri, 6 May 2005, Jesse Keating wrote:
> 
> > Any chance you could get like 3 so we could build a software RAID
> > array on them to see how fabulously they fail?  (:  What does Ed
> > think about using them in an array for data redundancy?
> 
> I plan to create a mixed md device with one aoe disk and whatever else
> I have on hand. But I have future designs on getting them to send one
> of their "shelves" (10-pack O' "blades") depending on how things go
> with the single kit.
> 
> How badly would the test be flawed if it was just one aoe device in a
> RAID? I mean, at least for seeing how/if it simply locks up like plain
> IDE's under md currently do (do they still? It's been two years since
> I did any real testing).

You might want to try it by itself first, no raid, to see if the disk
recovers from loss of network and loss of power to the disk ( looks
easy to pull the power ). Then setup a software raid 1 array and see
what happens. You might also try it with a firewire / usb
disk if this fails to be sure you're not bound by the ide interface. 

Bonnie++ results would be interesting on the "by itself" setup.
Comparing that run to the same disk ( I assume it's not too hard to
install and remove the disk from the blade ) attached to the
onboard ide1 of the same computer would be interesting. It would
also be interesting to see a run with a crossover cable and then with a
switch. A run on a non-managed switch and a managed switch or
another manufactures non-managed switch to see what effects a switch can
have on throughput. 

If you get a managed switch that can do port mirroring then a laptop
with Ethereal to watch what happens when the disks fails; power or
pulling the Ethernet from switch to blade, and recovers would be very
interesting. Ethereal has an AOE decoder and probably with the later
versions IO and other statistics. Maybe a hub if a switch with port
mirroring can't be found. I would be interest in how much broadcast
traffic is caused by failures and recouveries. 

Stats on the traffic when the blade is maxed would be
interesting, is the control traffic impacted / broken by heavy use?
Two blades would be interesting in this case.

Setup GFS and two nodes sharing the disk. Maybe setup mysql on the blade
and run the database benchmarks from both nodes?

I'm sure there are other interesting ways to play with the blade, but
that's all I can think of at the moment. Sound like a lot of fun no
matter how it turns out.

Thank you
Russell