HAST (Highly Available Storage) now in HEAD.

[Garrett Cooper]

On Fri, Feb 19, 2010 at 3:37 PM, Pawel Jakub Dawidek <pjd at freebsd.org> wrote:
> On Fri, Feb 19, 2010 at 03:11:44PM -0800, Garrett Cooper wrote:
>>     Very cool stuff. How many nodes max are you targeting for this
>> service [...]
>
> Currently HAST is intended for use only with High Availability clusters,
> not for performance clusters and is limited to exactly two nodes - one
> primary node, which has access to shared storage and one secondary node,
> which just receives updates from primary. User's applications should
> only work on primary node.

I was looking forward a bit more than what's setup today TBH, but that's ok...

>> [...] and what are some of the performance numbers for syncing
>> across the network (say with a 1GigE or 10GigE connection)?
>
> HAST should be able to saturate 1GigE link if you have fast enough
> storage. I've patches in the works to save data copying between userland
> and kernel. Currently, eg. write I/O request comes from the kernel, it
> is copied to hastd userland daemon, hastd copies it back to the kernel
> when writing to local component and then copies it again to the kernel
> when sending over network to secondary node. In other words a lot of
> copying. I prototyped a model where data is not copied at all between
> userland and kernel, but it needs a bit more work.

Ouch. Lots of CoW / copyout...? Is the data being checksummed /
verified somewhere to ensure integrity at all (I would think so / hope
so...)?
Cheers!
-Garrett