[tahoe-dev] Real-world Tahoe-LAFS grid deployment

[Francois Deppierraz]

Hi Nathan,

On 11/14/2010 05:41 PM, Nathan Eisenberg wrote:

> This is precisely why I'm hesitant to deploy multiple tahoe node
> processes on a single physical server.  That whole concept is a bit
> messy anyways - I'd much rather see 1 node process with 24 different
> storage threads (1 per disk) than 24 independent node processes.

What do you exactly mean by 'storage threads'? Do you see it as running
a single storage node on a RAID-0 array built on 24 disks?

I think that you lose the main advantage of erasure coding with this
approach. In theory, erasure coding should bring much better reliability
than full copies for a given expansion factor.

This also means that the same reliability can be achieved with less disk
space. This is something which is not really demonstrated by the default
3-of-10 encoding and its 3.3 expansion factor. Three full copies of a
given file is probably already considered reliable enough by most people
so what's the point of using erasure coding?

A quick computation on the Tahoe-LAFS Reliability Tool shows that
5-of-10 should have similar probability of loss with maintenance than
1-of-3 in a ten year period.

http://localhost:3456/reliability/?drive_lifetime=8Y&k=1&R=2&N=3&delta=1M&check_period=1M&report_period=3M&report_span=10Y
http://localhost:3456/reliability/?drive_lifetime=8Y&k=5&R=8&N=10&delta=1M&check_period=1M&report_period=3M&report_span=10Y

François