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.
> Does the ram disk actually set SWP_SOLIDSTATE?
Yes. "blk_queue_flag_set(QUEUE_FLAG_NONROT, q)" is called in
drivers/nvdimm/pmem.c.
Best Regards,
Huang, Ying
