"Huang, Ying" <ying.huang@xxxxxxxxx> writes:
> Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> writes:
>
>> On Wed, 20 May 2020 11:15:02 +0800 Huang Ying <ying.huang@xxxxxxxxx> wrote:
>>
>>> In some swap scalability test, it is found that there are heavy lock
>>> contention on swap cache even if we have split one swap cache radix
>>> tree per swap device to one swap cache radix tree every 64 MB trunk in
>>> commit 4b3ef9daa4fc ("mm/swap: split swap cache into 64MB trunks").
>>>
>>> The reason is as follow. After the swap device becomes fragmented so
>>> that there's no free swap cluster, the swap device will be scanned
>>> linearly to find the free swap slots. swap_info_struct->cluster_next
>>> is the next scanning base that is shared by all CPUs. So nearby free
>>> swap slots will be allocated for different CPUs. The probability for
>>> multiple CPUs to operate on the same 64 MB trunk is high. This causes
>>> the lock contention on the swap cache.
>>>
>>> To solve the issue, in this patch, for SSD swap device, a percpu
>>> version next scanning base (cluster_next_cpu) is added. Every CPU
>>> will use its own per-cpu next scanning base. And after finishing
>>> scanning a 64MB trunk, the per-cpu scanning base will be changed to
>>> the beginning of another randomly selected 64MB trunk. In this way,
>>> the probability for multiple CPUs to operate on the same 64 MB trunk
>>> is reduced greatly. Thus the lock contention is reduced too. For
>>> HDD, because sequential access is more important for IO performance,
>>> the original shared next scanning base is used.
>>>
>>> To test the patch, we have run 16-process pmbench memory benchmark on
>>> a 2-socket server machine with 48 cores. One ram disk is configured
>>
>> What does "ram disk" mean here? Which drivers(s) are in use and backed
>> by what sort of memory?
>
> We use the following kernel command line
>
> memmap=48G!6G memmap=48G!68G
>
> to create 2 DRAM based /dev/pmem disks (48GB each). Then we use these
> ram disks as swap devices.
>
>>> as the swap device per socket. The pmbench working-set size is much
>>> larger than the available memory so that swapping is triggered. The
>>> memory read/write ratio is 80/20 and the accessing pattern is random.
>>> In the original implementation, the lock contention on the swap cache
>>> is heavy. The perf profiling data of the lock contention code path is
>>> as following,
>>>
>>> _raw_spin_lock_irq.add_to_swap_cache.add_to_swap.shrink_page_list: 7.91
>>> _raw_spin_lock_irqsave.__remove_mapping.shrink_page_list: 7.11
>>> _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 2.51
>>> _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 1.66
>>> _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.29
>>> _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.03
>>> _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 0.93
>>>
>>> After applying this patch, it becomes,
>>>
>>> _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 3.58
>>> _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 2.3
>>> _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 2.26
>>> _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.8
>>> _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.19
>>>
>>> The lock contention on the swap cache is almost eliminated.
>>>
>>> And the pmbench score increases 18.5%. The swapin throughput
>>> increases 18.7% from 2.96 GB/s to 3.51 GB/s. While the swapout
>>> throughput increases 18.5% from 2.99 GB/s to 3.54 GB/s.
>>
>> If this was backed by plain old RAM, can we assume that the performance
>> improvement on SSD swap is still good?
>
> We need really fast disk to show the benefit. I have tried this on 2
> Intel P3600 NVMe disks. The performance improvement is only about 1%.
> The improvement should be better on the faster disks, such as Intel
> Optane disk. I will try to find some to test.
I finally find 2 Intel Optane disks to test. The pmbench throughput
(page accesses per second) increases ~1.7% with the patch. The swapin
throughput increases 2% (~1.36 GB/s to ~1.39 GB/s), the swapout
throughput increases 1.7% (~1.61 GB/s to 1.63 GB/s). Perf profile shows
the CPU cycles on the swap cache radix tree spinlock is reduced from
~1.76% to nearly 0. So the performance difference is much smaller, but
still measurable.
Best Regards,
Huang, Ying
