[RFC PATCH v3 0/2] scheduler: expose the topology of clusters and add cluster scheduler

[Dietmar Eggemann]

On 08/01/2021 22:30, Song Bao Hua (Barry Song) wrote:
>  
>> -----Original Message-----
>> From: Morten Rasmussen [mailto:morten.rasmussen at arm.com]
>> Sent: Saturday, January 9, 2021 4:13 AM
>> To: Tim Chen <tim.c.chen at linux.intel.com>
>> Cc: Song Bao Hua (Barry Song) <song.bao.hua at hisilicon.com>;
>> valentin.schneider at arm.com; catalin.marinas at arm.com; will at kernel.org;
>> rjw at rjwysocki.net; vincent.guittot at linaro.org; lenb at kernel.org;
>> gregkh at linuxfoundation.org; Jonathan Cameron <jonathan.cameron at huawei.com>;
>> mingo at redhat.com; peterz at infradead.org; juri.lelli at redhat.com;
>> dietmar.eggemann at arm.com; rostedt at goodmis.org; bsegall at google.com;
>> mgorman at suse.de; mark.rutland at arm.com; sudeep.holla at arm.com;
>> aubrey.li at linux.intel.com; linux-arm-kernel at lists.infradead.org;
>> linux-kernel at vger.kernel.org; linux-acpi at vger.kernel.org;
>> linuxarm at openeuler.org; xuwei (O) <xuwei5 at huawei.com>; Zengtao (B)
>> <prime.zeng at hisilicon.com>; tiantao (H) <tiantao6 at hisilicon.com>
>> Subject: Re: [RFC PATCH v3 0/2] scheduler: expose the topology of clusters and
>> add cluster scheduler
>>
>> On Thu, Jan 07, 2021 at 03:16:47PM -0800, Tim Chen wrote:
>>> On 1/6/21 12:30 AM, Barry Song wrote:
>>>> ARM64 server chip Kunpeng 920 has 6 clusters in each NUMA node, and each
>>>> cluster has 4 cpus. All clusters share L3 cache data while each cluster
>>>> has local L3 tag. On the other hand, each cluster will share some
>>>> internal system bus. This means cache is much more affine inside one cluster
>>>> than across clusters.
>>>
>>> There is a similar need for clustering in x86.  Some x86 cores could share
>> L2 caches that
>>> is similar to the cluster in Kupeng 920 (e.g. on Jacobsville there are 6 clusters
>>> of 4 Atom cores, each cluster sharing a separate L2, and 24 cores sharing
>> L3).
>>> Having a sched domain at the L2 cluster helps spread load among
>>> L2 domains.  This will reduce L2 cache contention and help with
>>> performance for low to moderate load scenarios.
>>
>> IIUC, you are arguing for the exact opposite behaviour, i.e. balancing
>> between L2 caches while Barry is after consolidating tasks within the
>> boundaries of a L3 tag cache. One helps cache utilization, the other
>> communication latency between tasks. Am I missing something?
> 
> Morten, this is not true.
> 
> we are both actually looking for the same behavior. My patch also
> has done the exact same behavior of spreading with Tim's patch.

That's the case for the load-balance path because of the extra Sched
Domain (SD) (CLS/MC_L2) below MC.

But in wakeup you add code which leads to a different packing strategy.

It looks like that Tim's workload (SPECrate mcf) shows a performance
boost solely because of the changes the additional MC_L2 SD introduces
in load balance. The wakeup path is unchanged, i.e. llc-packing. IMHO we
have to carefully distinguish between packing vs. spreading in wakeup
and load-balance here.

> Considering the below two cases:
> Case 1. we have two tasks without any relationship running in a system with 2 clusters and 8 cpus.
> 
> Without the sched_domain of cluster, these two tasks might be put as below:
> +-------------------+            +-----------------+
> | +----+   +----+   |            |                 |
> | |task|   |task|   |            |                 |
> | |1   |   |2   |   |            |                 |
> | +----+   +----+   |            |                 |
> |                   |            |                 |
> |       cluster1    |            |     cluster2    |
> +-------------------+            +-----------------+
> With the sched_domain of cluster, load balance will spread them as below:
> +-------------------+            +-----------------+
> | +----+            |            | +----+          |
> | |task|            |            | |task|          |
> | |1   |            |            | |2   |          |
> | +----+            |            | +----+          |
> |                   |            |                 |
> |       cluster1    |            |     cluster2    |
> +-------------------+            +-----------------+
> 
> Then task1 and tasks2 get more cache and decrease cache contention.
> They will get better performance.
> 
> That is what my original patch also can make. And tim's patch
> is also doing. Once we add a sched_domain, load balance will
> get involved.
> 
> 
> Case 2. we have 8 tasks, running in a system with 2 clusters and 8 cpus.
> But they are working in 4 groups:
> Task1 wakes up task4
> Task2 wakes up task5
> Task3 wakes up task6
> Task4 wakes up task7
> 
> With my changing in select_idle_sibling, the WAKE_AFFINE mechanism will
> try to put task1 and 4, task2 and 5, task3 and 6, task4 and 7 in same clusters rather
> than putting all of them in the random one of the 8 cpus. However, the 8 tasks
> are still spreading among the 8 cpus with my change in select_idle_sibling
> as load balance is still working.
> 
> +---------------------------+    +----------------------+
> | +----+        +-----+     |    | +----+      +-----+  |
> | |task|        |task |     |    | |task|      |task |  |
> | |1   |        | 4   |     |    | |2   |      |5    |  |
> | +----+        +-----+     |    | +----+      +-----+  |
> |                           |    |                      |
> |       cluster1            |    |     cluster2         |
> |                           |    |                      |
> |                           |    |                      |
> | +-----+       +------+    |    | +-----+     +------+ |
> | |task |       | task |    |    | |task |     |task  | |
> | |3    |       |  6   |    |    | |4    |     |8     | |
> | +-----+       +------+    |    | +-----+     +------+ |
> +---------------------------+    +----------------------+

Your use-case (#tasks, runtime/period) seems to be perfectly crafted to
show the benefit of your patch on your specific system (cluster-size =
4). IMHO, this extra infrastructure especially in the wakeup path should
show benefits over a range of different benchmarks.

> Let's consider the 3rd case, that one would be more tricky:
> 
> task1 and task2 have close relationship and they are waker-wakee pair.
> With my current patch, select_idle_sidling() wants to put them in one
> cluster, load balance wants to put them in two clusters. Load balance will win. 
> Then maybe we need some same mechanism like adjusting numa imbalance:
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/kernel/sched/fair.c?id=b396f52326de20
> if we permit a light imbalance between clusters, select_idle_sidling()
> will win. And task1 and task2 get better cache affinity.

This would look weird to allow this kind of imbalance on CLS (MC_L2) and
NUMA domains but not on the MC domain for example.