[linux-lvm] Thin snapshot caching behaviour

[dw+]

Hi there,

While playing with LVM thin provisioning, I've noticed that snapshots
seem to have different caching semantics compared to their original thin
LV.

I've hunted everywhere for documentation that describes this difference,
or even an indication of on what layer it occurs, but I can find none.
Perhaps someone here could shed some light?

Thin volume behaves like a regular drive:

# echo 1 > /proc/sys/vm/drop_caches

# dd if=/dev/vg0/tv0 of=/dev/null bs=64k count=1000
65536000 bytes (66 MB) copied, 0.946389 s, 69.2 MB/s

# dd if=/dev/vg0/tv0 of=/dev/null bs=64k count=1000
65536000 bytes (66 MB) copied, 0.00810655 s, 8.1 GB/s

Create activated snapshot:

# lvcreate -kn -s -n tv0s /dev/vg0/tv0
Logical volume "tv0s" created

# dd if=/dev/vg0/tv0s of=/dev/null bs=64k count=1000
65536000 bytes (66 MB) copied, 1.00061 s, 65.5 MB/s

Second dd shows no/very little speedup:

# time dd if=/dev/vg0/tv0s of=/dev/null bs=64k count=1000
65536000 bytes (66 MB) copied, 0.921402 s, 71.1 MB/s

Thanks,


David

[Lars Ellenberg]

Post by dw+
Hi there,
While playing with LVM thin provisioning, I've noticed that snapshots
seem to have different caching semantics compared to their original thin
LV.
I've hunted everywhere for documentation that describes this difference,
or even an indication of on what layer it occurs, but I can find none.
Perhaps someone here could shed some light?
# echo 1 > /proc/sys/vm/drop_caches
# dd if=/dev/vg0/tv0 of=/dev/null bs=64k count=1000
65536000 bytes (66 MB) copied, 0.946389 s, 69.2 MB/s
# dd if=/dev/vg0/tv0 of=/dev/null bs=64k count=1000
65536000 bytes (66 MB) copied, 0.00810655 s, 8.1 GB/s
# lvcreate -kn -s -n tv0s /dev/vg0/tv0
Logical volume "tv0s" created
# dd if=/dev/vg0/tv0s of=/dev/null bs=64k count=1000
65536000 bytes (66 MB) copied, 1.00061 s, 65.5 MB/s
# time dd if=/dev/vg0/tv0s of=/dev/null bs=64k count=1000
65536000 bytes (66 MB) copied, 0.921402 s, 71.1 MB/s

# echo 3 > /proc/sys/vm/drop_caches
# free -m
# exec 77< /dev/vg0/tv0s
# time dd if=/dev/vg0/tv0s of=/dev/null bs=64k count=1000
# free -m
# time dd if=/dev/vg0/tv0s of=/dev/null bs=64k count=1000
# free -m
# exec 77<&-
# free -m


Background:
linux "forgets" the buffer cache for a block device,
if the last opener is gone (no one has it open anymore).

The "main" device has some internal references,
or is mounted... someone has it open, anyways.

At least that's my best guess as to what you are seeing there.