Re: [PATCH v6 1/3] perf/core: use rb trees for pinned/flexible groups
From: Alexey Budankov
Date: Mon Aug 07 2017 - 11:32:27 EST
On 07.08.2017 12:13, Peter Zijlstra wrote:
> On Mon, Aug 07, 2017 at 10:39:13AM +0200, Peter Zijlstra wrote:
>> On Mon, Aug 07, 2017 at 10:17:46AM +0300, Alexey Budankov wrote:
>>> Makes sense. The implementation becomes a bit simpler. The drawbacks
>>> may be several rotations of potentially big tree on the critical path,
>>> instead of updating four pointers in case of the tree of lists.
>> Yes, but like said, it allows implementing a better scheduler than RR,
>> allowing us to fix rotation artifacts where task runtimes are near the
>> rotation window.
Could you elaborate more on the artifacts or my be share some link to the theory?
>> A slightly more complicated, but also interested scheduling problem is
>> the per-cpu flexible vs the per-task flexible. Ideally we'd rotate them
>> at the same priority based on service, without strictly prioritizing the
>> per-cpu events.
>> Again, that is something that should be possible once we have a more
>> capable event scheduler.
>> So yes, cons and pros.. :-)
> Also, I think for AVL tree you could do the erase and (re)insert
> combined and then rebalance in one go, not sure RB allows the same
> thing, but it might be fun looking into.
Not sure if AVL is more practical here. You get better balancing what gives
you faster average search for the price of longer modifications
so yes, need to measure and compare ... :-)