On Mon, 25 Jul 2016 10:39:25 -0400 Kyle Walker <kwal...@redhat.com> wrote:
> Java workloads using the MappedByteBuffer library result in the fadvise()
> and madvise() syscalls being used extensively. Following recent readahead
> limiting alterations, such as 600e19af ("mm: use only per-device readahead
> limit") and 6d2be915 ("mm/readahead.c: fix readahead failure for
> memoryless NUMA nodes and limit readahead pages"), application performance
> suffers in instances where small readahead is configured.
Can this suffering be quantified please?
> By moving this limit outside of the syscall codepaths, the syscalls are
> able to advise an inordinately large amount of readahead when desired.
> With a cap being imposed based on the half of NR_INACTIVE_FILE and
> NR_FREE_PAGES. In essence, allowing performance tuning efforts to define a
> small readahead limit, but then benefiting from large sequential readahead
> values selectively.
>
> ...
>
> --- a/mm/readahead.c
> +++ b/mm/readahead.c
> @@ -211,7 +211,9 @@ int force_page_cache_readahead(struct address_space
> *mapping, struct file *filp,
> if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
> return -EINVAL;
>
> - nr_to_read = min(nr_to_read, inode_to_bdi(mapping->host)->ra_pages);
> + nr_to_read = min(nr_to_read, (global_page_state(NR_INACTIVE_FILE) +
> + (global_page_state(NR_FREE_PAGES)) / 2));
> +
> while (nr_to_read) {
> int err;
>
> @@ -484,6 +486,7 @@ void page_cache_sync_readahead(struct address_space
> *mapping,
>
> /* be dumb */
> if (filp && (filp->f_mode & FMODE_RANDOM)) {
> + req_size = min(req_size, inode_to_bdi(mapping->host)->ra_pages);
> force_page_cache_readahead(mapping, filp, offset, req_size);
> return;
> }
Linus probably has opinions ;)
