On Sat, 27 Jun 2020 04:13:04 +0100 Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote: > On Fri, Jun 26, 2020 at 02:23:03PM -0700, Tim Chen wrote: > > Enlarge the pagevec size to 31 to reduce LRU lock contention for > > large systems. > > > > The LRU lock contention is reduced from 8.9% of total CPU cycles > > to 2.2% of CPU cyles. And the pmbench throughput increases > > from 88.8 Mpages/sec to 95.1 Mpages/sec. > > The downside here is that pagevecs are often stored on the stack (eg > truncate_inode_pages_range()) as well as being used for the LRU list. > On a 64-bit system, this increases the stack usage from 128 to 256 bytes > for this array. > > I wonder if we could do something where we transform the ones on the > stack to DECLARE_STACK_PAGEVEC(pvec), and similarly DECLARE_LRU_PAGEVEC > the ones used for the LRUs. There's plenty of space in the header to > add an unsigned char sz, delete PAGEVEC_SIZE and make it an variable > length struct. > > Or maybe our stacks are now big enough that we just don't care. > What do you think? And I wonder how useful CONFIG_NR_CPUS is for making this decision. Presumably a lot of general-purpose kernel builds have CONFIG_NR_CPUS much larger than the actual number of CPUs. I can't think of much of a fix for this, apart from making it larger on all kernels, Is there a downside to this?
