On 2020/05/13 16:07, Hannes Reinecke wrote: > Hi all, > > here are two patches to improve performance when a regular device > is specified for dm-zoned. > Damien noted that we have a performance drop after the zones on the > regular device are used up, as then the random zones on the zoned > device are being used, _and_ reclaim is running on the same disk. > This slows down I/O whle using random zones, increasing again after > the random zones are used up. > This patchset fixes this by not allowing random zones for caching > if a regular device is present. > > Patch is relative to the dm-5.8 branch from > git://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm.git > > As usual, comments and reviews are welcome. Ran this for 2 hours, result plot attached. Same workload as before, 2 hours run. This looks much better. No drop to near nothing as there is now no HDD-to-HDD reclaim. The HDD is exercised at drive speed (150MB/s, which is about what I get with ext4 on a similar regular drive without dm-zoned). If anything, it looks like reclaim is starting too early. The initial peak performance at SSD speed is now very very short. We need to check the dmz_should_reclaim() changes. > > > Hannes Reinecke (2): > dm-zoned: invert zone check in dmz_reset_zone() > dm-zoned: split off random and cache zones > > .../admin-guide/device-mapper/dm-zoned.rst | 17 +- > drivers/md/dm-zoned-metadata.c | 157 +++++++++++++----- > drivers/md/dm-zoned-reclaim.c | 70 ++++---- > drivers/md/dm-zoned-target.c | 19 ++- > drivers/md/dm-zoned.h | 7 +- > 5 files changed, 189 insertions(+), 81 deletions(-) > -- Damien Le Moal Western Digital Research
Attachment:
dm-zoned-cache.png
Description: dm-zoned-cache.png
-- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
