Messages in this thread | | | Date | Wed, 23 Dec 2009 09:49:35 +0100 | From | Ingo Molnar <> | Subject | Re: workqueue thing |
| |
* Tejun Heo <tj@kernel.org> wrote:
> Hello, > > On 12/23/2009 05:37 PM, Ingo Molnar wrote: > >> Sure, fair enough but there's also a different side. It'll allow much > >> easier implementation of things like in-kernel media presence polling (I > >> have some code for this but it's still just forming) and per-device. It > >> gives a much easier tool to extract concurrency and thus opens up new > >> possibilities. > >> > >> So, anyways, alright, I'll go try some conversions. > > > > Well, but note that you are again talking performance. Concurrency > > _IS_ performance: either in terms of reduced IO/app/request latency > > or in terms of CPU utilization. > > I wasn't talking about performance above. Easiness or flexibility to > extract concurrency opens up possibilities for new things or easier ways of > doing things. It affects the design process. You don't have to jump > through hoops for concurrency management and removing that restriction > results in lower amount of convolution and simplifies design.
Which is why i said this in the next paragraph:
> > ( Plus reduction in driver complexity can be measured as well, in the > > diffstat space.)
A new facility that is so mysterious that it cannot be shown to have any performance/scalability/latency benefit _nor_ can it be shown to reduce driver complexity simply does not exist IMO.
A tangle benefit has to show up _somewhere_ - if not in the performance space then in the diffstat space (and vice versa) - that's all what i'm arguing.
Ingo
| |