Hi Andrew,
On Mon, Jun 12, 2023 at 02:43:31PM -0700, Andrew Morton wrote:
> On Fri, 9 Jun 2023 23:07:30 -0400 Yu Ma <yu.ma@xxxxxxxxx> wrote:
>
> > When running UnixBench/Execl throughput case, false sharing is observed
> > due to frequent read on base_addr and write on free_bytes, chunk_md.
> >
> > UnixBench/Execl represents a class of workload where bash scripts
> > are spawned frequently to do some short jobs. It will do system call on
> > execl frequently, and execl will call mm_init to initialize mm_struct
> > of the process. mm_init will call __percpu_counter_init for
> > percpu_counters initialization. Then pcpu_alloc is called to read
> > the base_addr of pcpu_chunk for memory allocation. Inside pcpu_alloc,
> > it will call pcpu_alloc_area to allocate memory from a specified chunk.
> > This function will update "free_bytes" and "chunk_md" to record the
> > rest free bytes and other meta data for this chunk. Correspondingly,
> > pcpu_free_area will also update these 2 members when free memory.
> > Call trace from perf is as below:
> > + 57.15% 0.01% execl [kernel.kallsyms] [k] __percpu_counter_init
> > + 57.13% 0.91% execl [kernel.kallsyms] [k] pcpu_alloc
> > - 55.27% 54.51% execl [kernel.kallsyms] [k] osq_lock
> > - 53.54% 0x654278696e552f34
> > main
> > __execve
> > entry_SYSCALL_64_after_hwframe
> > do_syscall_64
> > __x64_sys_execve
> > do_execveat_common.isra.47
> > alloc_bprm
> > mm_init
> > __percpu_counter_init
> > pcpu_alloc
> > - __mutex_lock.isra.17
> >
> > In current pcpu_chunk layout, ‘base_addr’ is in the same cache line
> > with ‘free_bytes’ and ‘chunk_md’, and ‘base_addr’ is at the
> > last 8 bytes. This patch moves ‘bound_map’ up to ‘base_addr’,
> > to let ‘base_addr’ locate in a new cacheline.
> >
> > With this change, on Intel Sapphire Rapids 112C/224T platform,
> > based on v6.4-rc4, the 160 parallel score improves by 24%.
>
> Well that's nice.
>
> >
> > ...
> >
> > --- a/mm/percpu-internal.h
> > +++ b/mm/percpu-internal.h
> > @@ -41,10 +41,17 @@ struct pcpu_chunk {
> > struct list_head list; /* linked to pcpu_slot lists */
> > int free_bytes; /* free bytes in the chunk */
> > struct pcpu_block_md chunk_md;
> > - void *base_addr; /* base address of this chunk */
> > + unsigned long *bound_map; /* boundary map */
> > +
> > + /*
> > + * base_addr is the base address of this chunk.
> > + * To reduce false sharing, current layout is optimized to make sure
> > + * base_addr locate in the different cacheline with free_bytes and
> > + * chunk_md.
> > + */
> > + void *base_addr ____cacheline_aligned_in_smp;
> >
> > unsigned long *alloc_map; /* allocation map */
> > - unsigned long *bound_map; /* boundary map */
> > struct pcpu_block_md *md_blocks; /* metadata blocks */
> >
> > void *data; /* chunk data */
>
> This will of course consume more memory. Do we have a feel for the
> worst-case impact of this?
>
The pcpu_chunk struct is a backing data structure per chunk, so the
additional memory should not be dramatic. A chunk covers ballpark
between 64kb and 512kb memory depending on some config and boot time
stuff, so I believe the additional memory used here is nominal at best.
Working the #s on my desktop:
Percpu: 58624 kB
28 cores -> ~2.1MB of percpu memory.
At say ~128KB per chunk -> 33 chunks, generously 40 chunks.
Adding alignment might bump the chunk size ~64 bytes, so in total ~2KB
of overhead?
I believe we can do a little better to avoid eating that full padding,
so likely less than that.
Acked-by: Dennis Zhou <dennis@xxxxxxxxxx>
Thanks,
Dennis
