On Thu, Jul 20, 2023 at 11:05:17PM +0800, Hyeonggon Yoo wrote:
> > > > let me introduce our test process.
> > > >
> > > > we make sure the tests upon commit and its parent have exact same environment
> > > > except the kernel difference, and we also make sure the config to build the
> > > > commit and its parent are identical.
> > > >
> > > > we run tests for one commit at least 6 times to make sure the data is stable.
> > > >
> > > > such like for this case, we rebuild the commit and its parent's kernel, the
> > > > config is attached FYI.
> > >
> > > Hello Oliver,
> > >
> > > Thank you for confirming the testing environment is totally fine.
> > > and I'm sorry. I didn't mean to offend that your tests were bad.
> > >
> > > It was more like "oh, the data totally doesn't make sense to me"
> > > and I blamed the tests rather than my poor understanding of the data ;)
> > >
> > > Anyway,
> > > as the data shows a repeatable regression,
> > > let's think more about the possible scenario:
> > >
> > > I can't stop thinking that the patch must've affected the system's
> > > reclamation behavior in some way.
> > > (I think more active anon pages with a similar number total of anon
> > > pages implies the kernel scanned more pages)
> > >
> > > It might be because kswapd was more frequently woken up (possible if
> > > skbs were allocated with GFP_ATOMIC)
> > > But the data provided is not enough to support this argument.
> > >
> > > > 2.43 ± 7% +4.5 6.90 ± 11% perf-profile.children.cycles-pp.get_partial_node
> > > > 3.23 ± 5% +4.5 7.77 ± 9% perf-profile.children.cycles-pp.___slab_alloc
> > > > 7.51 ± 2% +4.6 12.11 ± 5% perf-profile.children.cycles-pp.kmalloc_reserve
> > > > 6.94 ± 2% +4.7 11.62 ± 6% perf-profile.children.cycles-pp.__kmalloc_node_track_caller
> > > > 6.46 ± 2% +4.8 11.22 ± 6% perf-profile.children.cycles-pp.__kmem_cache_alloc_node
> > > > 8.48 ± 4% +7.9 16.42 ± 8% perf-profile.children.cycles-pp._raw_spin_lock_irqsave
> > > > 6.12 ± 6% +8.6 14.74 ± 9% perf-profile.children.cycles-pp.native_queued_spin_lock_slowpath
> > >
> > > And this increased cycles in the SLUB slowpath implies that the actual
> > > number of objects available in
> > > the per cpu partial list has been decreased, possibly because of
> > > inaccuracy in the heuristic?
> > > (cuz the assumption that slabs cached per are half-filled, and that
> > > slabs' order is s->oo)
> >
> > From the patch:
> >
> > static unsigned int slub_max_order =
> > - IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : PAGE_ALLOC_COSTLY_ORDER;
> > + IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : 2;
> >
> > Could this be related? that it reduces the order for some slab cache,
> > so each per-cpu slab will has less objects, which makes the contention
> > for per-node spinlock 'list_lock' more severe when the slab allocation
> > is under pressure from many concurrent threads.
>
> hackbench uses skbuff_head_cache intensively. So we need to check if
> skbuff_head_cache's
> order was increased or decreased. On my desktop skbuff_head_cache's
> order is 1 and I roughly
> guessed it was increased, (but it's still worth checking in the testing env)
>
> But decreased slab order does not necessarily mean decreased number
> of cached objects per CPU, because when oo_order(s->oo) is smaller,
> then it caches
> more slabs into the per cpu slab list.
>
> I think more problematic situation is when oo_order(s->oo) is higher,
> because the heuristic
> in SLUB assumes that each slab has order of oo_order(s->oo) and it's
> half-filled. if it allocates
> slabs with order lower than oo_order(s->oo), the number of cached
> objects per CPU
> decreases drastically due to the inaccurate assumption.
>
> So yeah, decreased number of cached objects per CPU could be the cause
> of the regression due to the heuristic.
>
> And I have another theory: it allocated high order slabs from remote node
> even if there are slabs with lower order in the local node.
>
> ofc we need further experiment, but I think both improving the
> accuracy of heuristic and
> avoiding allocating high order slabs from remote nodes would make SLUB
> more robust.
I run the reproduce command in a local 2-socket box:
"/usr/bin/hackbench" "-g" "128" "-f" "20" "--process" "-l" "30000" "-s" "100"
And found 2 kmem_cache has been boost: 'kmalloc-cg-512' and
'skbuff_head_cache'. Only order of 'kmalloc-cg-512' was reduced
from 3 to 2 with the patch, while its 'cpu_partial_slabs' was bumped
from 2 to 4. The setting of 'skbuff_head_cache' was kept unchanged.
And this compiled with the perf-profile info from 0Day's report, that the
'list_lock' contention is increased with the patch:
13.71% 13.70% [kernel.kallsyms] [k] native_queued_spin_lock_slowpath - -
5.80% native_queued_spin_lock_slowpath;_raw_spin_lock_irqsave;__unfreeze_partials;skb_release_data;consume_skb;unix_stream_read_generic;unix_stream_recvmsg;sock_recvmsg;sock_read_iter;vfs_read;ksys_read;do_syscall_64;entry_SYSCALL_64_after_hwframe;__libc_read
5.56% native_queued_spin_lock_slowpath;_raw_spin_lock_irqsave;get_partial_node.part.0;___slab_alloc.constprop.0;__kmem_cache_alloc_node;__kmalloc_node_track_caller;kmalloc_reserve;__alloc_skb;alloc_skb_with_frags;sock_alloc_send_pskb;unix_stream_sendmsg;sock_write_iter;vfs_write;ksys_write;do_syscall_64;entry_SYSCALL_64_after_hwframe;__libc_write
Also I tried to restore the slub_max_order to 3, and the regression was
gone.
static unsigned int slub_max_order =
- IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : 2;
+ IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : 3;
static unsigned int slub_min_objects;
Thanks,
Feng
> > I don't have direct data to backup it, and I can try some experiment.
>
> Thank you for taking time for experiment!
>
> Thanks,
> Hyeonggon
>
> > > > then retest on this test machine:
> > > > 128 threads 2 sockets Intel(R) Xeon(R) Gold 6338 CPU @ 2.00GHz (Ice Lake) with 256G memory
