[svlug] SSD was: Slides from the GoLUG meeting

Ivan Sergio Borgonovo mail at webthatworks.it
Fri Jun 3 03:18:34 PDT 2016



On 06/03/2016 10:34 AM, Rick Moen wrote:
> Quoting Ivan Sergio Borgonovo (mail at webthatworks.it):
>
>> oh it does.
>> https://www.kernel.org/doc/Documentation/block/cfq-iosched.txt
>>
>> "CFQ has some optimizations for SSDs and if it detects a non-rotational
>> media which can support higher queue depth (multiple requests at in
>> flight at a time), then it cuts down on idling of individual queues and
>> all the queues move to sync-noidle tree and only tree idle remains. This
>> tree idling provides isolation with buffered write queues on async tree."

> The point is, people who've studied this matter carefully say the Noop
> and Deadline schedulers are simply better for such media, less impairing
> of I/O, less wasteful of CPU cycles doing inapplicable things.  I see no
> reason to doubt them, or to think that CFQ with some functions disabled
> is better still.

> What I cannot understand is why you think the quoted text shows that the
> kernel can do a better job than you.  It doesn't show that at all, and
> also good reasons why it's a doubtful assertion stand on their merits
> irrespective of what that file says.

I think I've read it from "people who've studied this matter carefully".
Sorry I couldn't track down where I read it, but the memories I have 
about it, and they could be faulty were "you shouldn't care anymore".

I think deciding what's best is *now* more complicated than "scheduler A 
is simpler and it looks more adequate".
I didn't find many recent benchmark comparing cfq to deadline since 
there is a new scheduler that is attracting attentions (bfq) but reading

https://blog.pgaddict.com/posts/postgresql-io-schedulers-cfq-noop-deadline

it looks as it is at least debatable which scheduler to chose.

I'm just suggesting that if anyone really really need to know should do 
his homework again.

-- 
Ivan Sergio Borgonovo
http://www.webthatworks.it




More information about the svlug mailing list