> On Apr 19, 2017, at 3:02 AM, Jan Kara <jack@xxxxxxx> wrote: > > On Tue 18-04-17 16:04:17, Josef Bacik wrote: >> By default we set DCACHE_REFERENCED and I_REFERENCED on any dentry or >> inode we create. This is problematic as this means that it takes two >> trips through the LRU for any of these objects to be reclaimed, >> regardless of their actual lifetime. With enough pressure from these >> caches we can easily evict our working set from page cache with single >> use objects. So instead only set *REFERENCED if we've already been >> added to the LRU list. This means that we've been touched since the >> first time we were accessed, and so more likely to need to hang out in >> cache. >> >> To illustrate this issue I wrote the following scripts >> >> https://github.com/josefbacik/debug-scripts/tree/master/cache-pressure >> >> on my test box. It is a single socket 4 core CPU with 16gib of RAM and >> I tested on an Intel 2tib NVME drive. The cache-pressure.sh script >> creates a new file system and creates 2 6.5gib files in order to take up >> 13gib of the 16gib of ram with pagecache. Then it runs a test program >> that reads these 2 files in a loop, and keeps track of how often it has >> to read bytes for each loop. On an ideal system with no pressure we >> should have to read 0 bytes indefinitely. The second thing this script >> does is start a fs_mark job that creates a ton of 0 length files, >> putting pressure on the system with slab only allocations. On exit the >> script prints out how many bytes were read by the read-file program. >> The results are as follows >> >> Without patch: >> /mnt/btrfs-test/reads/file1: total read during loops 27262988288 >> /mnt/btrfs-test/reads/file2: total read during loops 27262976000 >> >> With patch: >> /mnt/btrfs-test/reads/file2: total read during loops 18640457728 >> /mnt/btrfs-test/reads/file1: total read during loops 9565376512 >> >> This patch results in a 50% reduction of the amount of pages evicted >> from our working set. >> >> Signed-off-by: Josef Bacik <jbacik@xxxxxx> > > To me this makes sense and I was myself wondering why we add dentries and > inodes to the lru list with referenced bit set in the past. I would just > note that the patch is changing the balance between dentry/inode cache > eviction and page cache eviction (so that pages have lower chance of being > evicted) and your test measures only the benefit of less pages being > evicted. So to be fair you should also complement it with a test where > dentries & inodes actually do get eventually reused and how that works out > in presence of memory pressure from use-once pages in the page cache. > Yup that's fair. I wrote up a different script to exercise the warm icache/dcache vs once use pages, the scripts are in the same place as my original script https://github.com/josefbacik/debug-scripts/tree/master/cache-pressure Basically create 3million empty nodes, create a 100gib sparse file, stat in a loop all the files while cat'ing the sparse file to /dev/null. There's a bcc script in there to count how many times the stat'ing programs allocate a new inode, and with and without my patch the test exited with those processes creating 0 nodes. This makes sense as we will create the files and they will not have REFERENCED, and then as soon as the files are read in again they will be marked with REFERENCED, so they become just as hard to remove as they were originally. Thanks, Josef
