Current sheepdog driver has a big write lock section. It harms write
performance significantly. This patch removes a needless big write
lock section with checking submitted requests can cause COW or
not. The write performance can be improved significantly.
In a case of 4k rand write by fio:
- before: 1144 IOPS
- after: 19074 IOPS
Signed-off-by: Hitoshi Mitake <mitake.hitoshi@xxxxxxxxxxxxx>
---
usr/bs_sheepdog.c | 14 +++++++++++++-
1 file changed, 13 insertions(+), 1 deletion(-)
diff --git a/usr/bs_sheepdog.c b/usr/bs_sheepdog.c
index 0838009..9b5b259 100644
--- a/usr/bs_sheepdog.c
+++ b/usr/bs_sheepdog.c
@@ -876,6 +876,7 @@ static int sd_io(struct sheepdog_access_info *ai, int write, char *buf, int len,
uint16_t flags = 0;
int need_update_inode = 0, need_reload_inode;
int nr_copies = ai->inode.nr_copies;
+ int need_write_lock, check_idx;
goto do_req;
@@ -891,7 +892,18 @@ reload_in_read_path:
pthread_rwlock_unlock(&ai->inode_lock);
do_req:
- if (write)
+ need_write_lock = 0;
+ vid = ai->inode.vdi_id;
+
+ for (check_idx = idx; check_idx < max; check_idx++) {
+ if (ai->inode.data_vdi_id[check_idx] == vid)
+ continue;
+
+ need_write_lock = 1;
+ break;
+ }
+
+ if (need_write_lock)
pthread_rwlock_wrlock(&ai->inode_lock);
else
pthread_rwlock_rdlock(&ai->inode_lock);
--
1.9.1
--
To unsubscribe from this list: send the line "unsubscribe stgt" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
