Re: [Nbd] [CFT][PATCH 0/4] nbd-module: merge of FLUSH/FUA and DISCARD series

[Paolo B. <pbo...@re...>]

On 09/13/2011 02:28 PM, Alex Bligh wrote:
>
>
> --On 13 September 2011 14:05:41 +0200 Paolo Bonzini<pbo...@pu...>
> wrote:
>
>>> I see you've got support for NBD_TIMEOUT in your patch set too.
>>> That might be better broken out.
>>
>> It's just ioctl_cmd_to_ascii.
>
> doh. That's what I get for reading too fast.
>
>>> I get the feeling patch 3/4 is going not going to get applied.
>>> I should really fix nbd-client so it doesn't need this.
>>
>> FWIW I've seen multiple reports of getting better performance _without_
>> NONROT, but I don't see anything particularly bad in that patch.  Flags
>> are just a hint, you can always ignore them.
>
> Indeed.
>
> The case where you want NONROT set is where the elevator of the server is
> likely to outperform the client, and there is very little latency between
> the two ends (i.e. client elevator merges gain nothing). Small client
> VMs are an example.
>
> The case where you want NONROT unset is where the client elevator is useful
> (e.g. big memory server) and where the server will heavily benefit from
> adjacent requests being merged, e.g. where there is a caching strategy
> which is heavily localised.
>
> Personally, I think NONROT unset is a better default, but then I haven't
> done lots of timings. I would be interested to know how we came upon
> the current default though.
>
>>> Out of interest, what kernel(s) have you compile-tested these
>>> against? I'd quite like to get them into my standalone module
>>> tree (which is far far easier to test against).
>>
>> 3.0.3
>
> OK - I think my tree is based on 2.6.32, so I may have a little
> work to do.
>

Ping?

Paolo