[PATCH v5] enhance NUMA affinity heuristic

[Thomas Monjalon]

I'm not comfortable with this patch.

First, there is no comment in the code which helps to understand the logic.
Second, I'm afraid changing the value of the per-core variable _socket_id
may have an impact on some applications.


16/02/2023 03:50, You, KaisenX:
> From: Burakov, Anatoly <anatoly.burakov at intel.com>
> > On 2/1/2023 12:20 PM, Kaisen You wrote:
> > > Trying to allocate memory on the first detected numa node has less
> > > chance to find some memory actually available rather than on the main
> > > lcore numa node (especially when the DPDK application is started only
> > > on one numa node).
> > >
> > > Fixes: 705356f0811f ("eal: simplify control thread creation")
> > > Fixes: bb0bd346d5c1 ("eal: suggest using --lcores option")
> > > Cc: stable at dpdk.org
> > >
> > > Signed-off-by: David Marchand <david.marchand at redhat.com>
> > > Signed-off-by: Kaisen You <kaisenx.you at intel.com>
> > > ---
> > > Changes since v4:
> > > - mod the patch title,
> > >
> > > Changes since v3:
> > > - add the assignment of socket_id in thread initialization,
> > >
> > > Changes since v2:
> > > - add uncommitted local change and fix compilation,
> > >
> > > Changes since v1:
> > > - accomodate for configurations with main lcore running on multiples
> > >    physical cores belonging to different numa,
> > > ---
> > >   lib/eal/common/eal_common_thread.c | 1 +
> > >   lib/eal/common/malloc_heap.c       | 4 ++++
> > >   2 files changed, 5 insertions(+)
> > >
> > > diff --git a/lib/eal/common/eal_common_thread.c
> > > b/lib/eal/common/eal_common_thread.c
> > > index 38d83a6885..21bff971f8 100644
> > > --- a/lib/eal/common/eal_common_thread.c
> > > +++ b/lib/eal/common/eal_common_thread.c
> > > @@ -251,6 +251,7 @@ static void *ctrl_thread_init(void *arg)
> > >   	void *routine_arg = params->arg;
> > >
> > >   	__rte_thread_init(rte_lcore_id(), cpuset);
> > > +	RTE_PER_LCORE(_socket_id) = SOCKET_ID_ANY;
> > >   	params->ret = rte_thread_set_affinity_by_id(rte_thread_self(),
> > cpuset);
> > >   	if (params->ret != 0) {
> > >   		__atomic_store_n(&params->ctrl_thread_status,
> > > diff --git a/lib/eal/common/malloc_heap.c
> > > b/lib/eal/common/malloc_heap.c index d7c410b786..3ee19aee15 100644
> > > --- a/lib/eal/common/malloc_heap.c
> > > +++ b/lib/eal/common/malloc_heap.c
> > > @@ -717,6 +717,10 @@ malloc_get_numa_socket(void)
> > >   			return socket_id;
> > >   	}
> > >
> > > +	socket_id = rte_lcore_to_socket_id(rte_get_main_lcore());
> > > +	if (socket_id != (unsigned int)SOCKET_ID_ANY)
> > > +		return socket_id;
> > > +
> > >   	return rte_socket_id_by_idx(0);
> > >   }
> > >
> > 
> > I may be lacking context, but I don't quite get the suggested change.
> >  From what I understand, the original has to do with assigning lcore cpusets in
> > such a way that an lcore ends up having two socket ID's (because it's been
> > assigned to CPU's on different sockets). Why is this allowed in the first place?
> > It seems like a user error to me, as it breaks many of the fundamental
> > assumptions DPDK makes.
> > 
> In a dual socket system, if all used cores are in socket 1 and the NIC is in socket 1, 
> no memory is allocated for socket 0. This is to optimize memory consumption.
> 
> I agree with you. If the startup parameters can ensure that both sockets 
> allocate memory, there will be no problem.
> However, due to the different CPU topologies of different systems, 
> It is difficult for users to ensure that the startup parameter contains two cpu nodes.
> 
> > I'm fine with using main lcore socket for control threads, I just don't think the
> > `socket_id != SOCKET_ID_ANY` thing should be checked here, because it
> > apparently tries to compensate for a problem with cpuset of the main thread,
> > which shouldn't have happened to begin with.
> > 
> This issue has been explained in detail in the discussion of the patch v1 version. 
> I will forward the previous email to you. The content of the email will also better 
> let you know the purpose of submitting this patch.