Re: [PATCH v3] sched/fair: Add advisory flag for borrowing a timeslice

[Rik van Riel <riel@redhat.com>]

On 11/25/2014 12:22 PM, Khalid Aziz wrote:
> On 11/25/2014 08:25 AM, Rik van Riel wrote:
>> On 11/25/2014 09:52 AM, Khalid Aziz wrote:
>>> On 11/24/2014 07:03 PM, Rik van Riel wrote:
>>>> -----BEGIN PGP SIGNED MESSAGE-----
>>>> Hash: SHA1
>>>>
>>>> On 11/24/2014 03:56 PM, Khalid Aziz wrote:
>>>>> sched/fair: Add advisory flag for borrowing a timeslice
>>>>>
>>>>> This patch adds a way for a task to request to borrow one
>>>>> timeslice from future if it is about to be preempted, so it could
>>>>> delay preemption and complete any critical task it is in the middle
>>>>> of.
>>>>>
>>>>> This feature helps with performance on databases and has been used
>>>>> for many years on other OSs by the databases. This feature helps in
>>>>> situation where a task acquires a lock before performing a critical
>>>>> operation on the database and happens to get preempted
>>>>
>>>> Why don't the other tasks that want the lock sleep on the
>>>> lock?
>>>>
>>>> I can see this "solution" help mostly with userspace spinlocks,
>>>> which are relics of a past era that need to die. There is no
>>>> way userspace spinlocks will not fail miserably on virtual
>>>> machines, and it is time to get rid of them.
>>>
>>> This solution indeed is for userspace spinlocks. Database code has been
>>> written with all critical locking implemented in userspace (as I have
>>> been told by database folks. I am not a database guy).
>>
>> They should fix that.
> 
> Agreed, but that is a very large task for databases for sure and
> potentially for java as well (think exhaustive testing as well besides
> code re-write). I do not see this happening except as part of major
> re-architecture. In the mean time, if we can give them help without
> impacting kernel significantly for other users, it only makes Linux a
> more usable platform for them.

I am not convinced that permanently adding overhead to the
Linux scheduler in order to deal with a temporary problem in
database and Java software (probably not every JVM, either!)
is a worthwhile trade-off..