-----Original Message----- From: Dietmar Eggemann [mailto:dietmar.eggemann@arm.com] Sent: Wednesday, January 13, 2021 12:00 AM To: Morten Rasmussen morten.rasmussen@arm.com; Tim Chen tim.c.chen@linux.intel.com Cc: Song Bao Hua (Barry Song) song.bao.hua@hisilicon.com; valentin.schneider@arm.com; catalin.marinas@arm.com; will@kernel.org; rjw@rjwysocki.net; vincent.guittot@linaro.org; lenb@kernel.org; gregkh@linuxfoundation.org; Jonathan Cameron jonathan.cameron@huawei.com; mingo@redhat.com; peterz@infradead.org; juri.lelli@redhat.com; rostedt@goodmis.org; bsegall@google.com; mgorman@suse.de; mark.rutland@arm.com; sudeep.holla@arm.com; aubrey.li@linux.intel.com; linux-arm-kernel@lists.infradead.org; linux-kernel@vger.kernel.org; linux-acpi@vger.kernel.org; linuxarm@openeuler.org; xuwei (O) xuwei5@huawei.com; Zengtao (B) prime.zeng@hisilicon.com; tiantao (H) tiantao6@hisilicon.com Subject: Re: [RFC PATCH v3 0/2] scheduler: expose the topology of clusters and add cluster scheduler
On 11/01/2021 10:28, Morten Rasmussen wrote:
On Fri, Jan 08, 2021 at 12:22:41PM -0800, Tim Chen wrote:
On 1/8/21 7:12 AM, Morten Rasmussen wrote:
On Thu, Jan 07, 2021 at 03:16:47PM -0800, Tim Chen wrote:
On 1/6/21 12:30 AM, Barry Song wrote:
[...]
I think it is going to depend on the workload. If there are dependent tasks that communicate with one another, putting them together in the same cluster will be the right thing to do to reduce communication costs. On the other hand, if the tasks are independent, putting them together
on the same cluster
will increase resource contention and spreading them out will be better.
Agree. That is exactly where I'm coming from. This is all about the task placement policy. We generally tend to spread tasks to avoid resource contention, SMT and caches, which seems to be what you are proposing to extend. I think that makes sense given it can produce significant benefits.
Any thoughts on what is the right clustering "tag" to use to clump related tasks together? Cgroup? Pid? Tasks with same mm?
I think this is the real question. I think the closest thing we have at the moment is the wakee/waker flip heuristic. This seems to be related. Perhaps the wake_affine tricks can serve as starting point?
wake_wide() switches between packing (select_idle_sibling(), llc_size CPUs) and spreading (find_idlest_cpu(), all CPUs).
AFAICS, since none of the sched domains set SD_BALANCE_WAKE, currently all wakeups are (llc-)packed.
select_task_rq_fair()
for_each_domain(cpu, tmp)
if (tmp->flags & sd_flag) sd = tmp;In case we would like to further distinguish between llc-packing and even narrower (cluster or MC-L2)-packing, we would introduce a 2. level packing vs. spreading heuristic further down in sis().
IMHO, Barry's current implementation doesn't do this right now. Instead he's trying to pack on cluster first and if not successful look further among the remaining llc CPUs for an idle CPU.
Right now in the main cases of using wake_affine to achieve better performance, processes are actually bound within one numa which is also a LLC in kunpeng920.
Probably LLC=NUMA is also true for X86 Jacobsville, Tim?
So one possible way to pretend a 2-level packing might be: if the affinity cpuset of waker and waker are both subset of one same LLC, we totally use cluster as the factor to determine packing or not and ignore LLC.
I haven't really done this, but the below code can make the same result by forcing llc_id=cluster_id:
diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c index d72eb8d..3d78097 100644 --- a/arch/arm64/kernel/topology.c +++ b/arch/arm64/kernel/topology.c @@ -107,7 +107,7 @@ int __init parse_acpi_topology(void) cpu_topology[cpu].cluster_id = topology_id; topology_id = find_acpi_cpu_topology_package(cpu); cpu_topology[cpu].package_id = topology_id; - +#if 0 i = acpi_find_last_cache_level(cpu);
if (i > 0) { @@ -119,8 +119,11 @@ int __init parse_acpi_topology(void) if (cache_id > 0) cpu_topology[cpu].llc_id = cache_id; } - } +#else + cpu_topology[cpu].llc_id = cpu_topology[cpu].cluster_id; +#endif
+ } return 0; } #endif
With this, I have seen some major improvement in hackbench especially for monogamous communication model (fds_num=1, one sender for one receiver): numactl -N 0 hackbench -p -T -l 200000 -f 1 -g $1
I have tested -g(group_nums) 6, 12, 18, 24, 28, 32, For each different g, I ran 20 times and got the average value. The result is as below:
g= 6 12 18 24 28 32 w/o 1.3243 1.6741 1.7560 1.9036 2.0262 2.1826 w/ 1.1314 1.1864 1.4494 1.6159 1.9078 2.1249
Using top -H and hit "f" to show cpu of each thread, I am seeing the two threads in one group are likely to run in a cluster. That's why the hackbench latency is decreasing much.
Thanks Barry