Re: [RFC 2/3] sched/topology: fix sched groups on NUMA machines with mesh topology

From: Lauro Venancio
Date: Thu Apr 13 2017 - 17:07:19 EST


On 04/13/2017 05:21 PM, Lauro Venancio wrote:
> On 04/13/2017 12:48 PM, Peter Zijlstra wrote:
>> On Thu, Apr 13, 2017 at 10:56:08AM -0300, Lauro Ramos Venancio wrote:
>>> Currently, on a 4 nodes NUMA machine with ring topology, two sched
>>> groups are generated for the last NUMA sched domain. One group has the
>>> CPUs from NUMA nodes 3, 0 and 1; the other group has the CPUs from nodes
>>> 1, 2 and 3. As CPUs from nodes 1 and 3 belongs to both groups, the
>>> scheduler is unable to directly move tasks between these nodes. In the
>>> worst scenario, when a set of tasks are bound to nodes 1 and 3, the
>>> performance is severely impacted because just one node is used while the
>>> other node remains idle.
>> I feel a picture would be ever so much clearer.
>>
>>> This patch constructs the sched groups from each CPU perspective. So, on
>>> a 4 nodes machine with ring topology, while nodes 0 and 2 keep the same
>>> groups as before [(3, 0, 1)(1, 2, 3)], nodes 1 and 3 have new groups
>>> [(0, 1, 2)(2, 3, 0)]. This allows moving tasks between any node 2-hops
>>> apart.
>> So I still have no idea what specifically goes wrong and how this fixes
>> it. Changelog is impenetrable.
> On a 4 nodes machine with ring topology, the last sched domain level
> contains groups with 3 numa nodes each. So we have four possible groups:
> (0, 1, 2) (1, 2, 3) (2, 3, 0)(3, 0, 1). As we need just two groups to
> fill the sched domain, currently, the groups (3, 0, 1) and (1, 2, 3) are
> used for all CPUs. The problem with it is that nodes 1 and 3 belongs to
> both groups, becoming impossible to move tasks between these two nodes.
>
> This patch uses different groups depending on the CPU they are
> installed. So nodes 0 and 2 CPUs keep the same group as before: (3, 0,
> 1) and (1, 2, 3). Nodes 1 and 3 CPUs use the new groups: (0, 1, 2) and
> (2, 3, 0). So the first pair of groups allows movement between nodes 0
> and 2; and the second pair of groups allows movement between nodes 1 and 3.
>
> I will improve the changelog.
>
>> "From each CPU's persepective" doesn't really help, there already is a
>> for_each_cpu() in.
> The for_each_cpu() is used to iterate across all sched domain cpus. It
> doesn't consider the CPU where the groups are being installed (parameter
> cpu in build_overlap_sched_groups()). Currently, the parameter cpu is
> used just for memory allocation and for ordering the groups, it doesn't
> change the groups that are chosen. This patch uses the parameter cpu to
> choose the first group, changing also, as consequence, the second group.
>> Also, since I'm not sure what happend to the 4 node system, I cannot
>> begin to imagine what would happen on the 8 node one.

Just for clarification, I am sending the nodes distance table for the
two most common typologies affected by this issue.

4 nodes, ring topology
node distances:
node 0 1 2 3
0: 10 20 30 20
1: 20 10 20 30
2: 30 20 10 20
3: 20 30 20 10

8 node, mesh topology
node distances:
node 0 1 2 3 4 5 6 7
0: 10 16 16 22 16 22 16 22
1: 16 10 16 22 22 16 22 16
2: 16 16 10 16 16 16 16 22
3: 22 22 16 10 16 16 22 16
4: 16 22 16 16 10 16 16 16
5: 22 16 16 16 16 10 22 22
6: 16 22 16 22 16 22 10 16
7: 22 16 22 16 16 22 16 10