It is worth noting that using dededuplication does consume a lot of memory.
Deduplication tables consume memory and eventually spill over and consume disk space - this causes extra read and write operations for every block of data on which deduplication is attempted cause more I/O waits.
A system with large pools with small memory areas does not perform deduplication well.
So, if you run the following command on your pool (pool-ssd in this example)
sudo zdb -s pool-ssd
..you see something like the following:
pool: pool-ssd
state: ONLINE
scan: scrub repaired 0B in 0 days 00:31:14 with 0 errors on Wed Sep 2 14:35:52 2020
config:
---
Looking at the "dedup: DDT entries 4258028, size 448B on disk, 306B in core" line, we have:
4258028 de-dup entries, each entry used 306 bytes, so that's ~1.2GB of memory for just the de-dup/ table. But it is capped at 1/4 the size of the ARC, so you may find on a small memory system such as a raspberry pi that a lot of the dedup table can't be stored in memory and needs to be on-disk, and this increases the I/O waits.
It may be an idea to disable de-dup and see if it helps to resolve your issue on a comparatively slow and memory constrained system such as a raspberry pi when you have large pools
It is worth noting that using dededuplication does consume a lot of memory.
Deduplication tables consume memory and eventually spill over and consume disk space - this causes extra read and write operations for every block of data on which deduplication is attempted cause more I/O waits.
A system with large pools with small memory areas does not perform deduplication well.
So, if you run the following command on your pool (pool-ssd in this example)
sudo zdb -s pool-ssd
..you see something like the following:
pool: pool-ssd
state: ONLINE
scan: scrub repaired 0B in 0 days 00:31:14 with 0 errors on Wed Sep 2 14:35:52 2020
config:
NAME STATE READ WRITE CKSUM INTEL_SSDSC2BW4 80H6_CVTR608104 NW480EGN ONLINE 0 0 0 INTEL_SSDSC2BW4 80H6_CVTR552500 S0480EGN ONLINE 0 0 0
pool-ssd ONLINE 0 0 0
mirror-0 ONLINE 0 0 0
ata-
ata-
errors: No known data errors
dedup: DDT entries 4258028, size 448B on disk, 306B in core
bucket allocated referenced _______ _______ _______ __ _______ _______ _______ _______ __
______ _______
refcnt blocks LSIZE PSIZE DSIZE blocks LSIZE PSIZE DSIZE
------ ------ ----- ----- ----- ------ ----- ----- -----
1 3.63M 464G 226G 226G 3.63M 464G 226G 226G
2 376K 46.1G 24.3G 24.3G 822K 100G 53.2G 53.2G
4 51.3K 5.60G 2.51G 2.51G 240K 26.3G 11.8G 11.8G
8 7.49K 925M 544M 544M 77.9K 9.41G 5.47G 5.47G
16 3.32K 415M 261M 261M 68.8K 8.41G 5.35G 5.35G
32 657 78.7M 62.1M 62.1M 29.3K 3.51G 2.79G 2.79G
64 389 46.2M 42.7M 42.7M 34.0K 4.03G 3.78G 3.78G
128 248 30.1M 29.5M 29.5M 44.8K 5.44G 5.32G 5.32G
256 339 42.1M 40.8M 40.8M 123K 15.4G 14.9G 14.9G
512 374 46.8M 46.1M 46.1M 271K 33.9G 33.5G 33.5G
1K 254 31.8M 31.3M 31.3M 355K 44.4G 43.8G 43.8G
2K 5 640K 513K 513K 12.9K 1.61G 1.20G 1.20G
4K 4 512K 512K 512K 22.7K 2.84G 2.84G 2.84G
8K 8 1M 897K 897K 89.3K 11.2G 10.1G 10.1G
64K 2 256K 6K 6K 226K 28.3G 679M 679M
Total 4.06M 517G 254G 254G 5.99M 759G 421G 421G
---
Looking at the "dedup: DDT entries 4258028, size 448B on disk, 306B in core" line, we have:
4258028 de-dup entries, each entry used 306 bytes, so that's ~1.2GB of memory for just the de-dup/ table. But it is capped at 1/4 the size of the ARC, so you may find on a small memory system such as a raspberry pi that a lot of the dedup table can't be stored in memory and needs to be on-disk, and this increases the I/O waits.
It may be an idea to disable de-dup and see if it helps to resolve your issue on a comparatively slow and memory constrained system such as a raspberry pi when you have large pools