Hi
Recently I just newly installed a new FreeBSD 10.1-STABLE server and discovered that 1MB recordsize support has been added. Theoretically, large recordsize will wasting more space, increase io latency especially reading a 4K data from 1MB block.
However with LZ4 compression, accessing to disk is minimal when reading a 4K data on 1MB recordsize since the compression ratio on the empty data is high, so that only 4K or less physical block on disk been accessed.
The overhead on writing 4k data to disk also minimal because the latency of inline compression is very low on high performance CPU .
May I know whether my assumption are correct? If so enabling 1MB recordsize won't harm the RW performance on ZFS, in addition, it also increases the compression ratio as well and speed up scrubbing process.
Recently I just newly installed a new FreeBSD 10.1-STABLE server and discovered that 1MB recordsize support has been added. Theoretically, large recordsize will wasting more space, increase io latency especially reading a 4K data from 1MB block.
However with LZ4 compression, accessing to disk is minimal when reading a 4K data on 1MB recordsize since the compression ratio on the empty data is high, so that only 4K or less physical block on disk been accessed.
The overhead on writing 4k data to disk also minimal because the latency of inline compression is very low on high performance CPU .
May I know whether my assumption are correct? If so enabling 1MB recordsize won't harm the RW performance on ZFS, in addition, it also increases the compression ratio as well and speed up scrubbing process.
On real data compression will be less efficient, and as result to read 4KB your storage may need to read 1MB. For write situation is even worse, because most of writes will cause read-modify-write cycle of 1MB block. Large blocks indeed improve compression and reduce processing overhead for large files, but, unless you have very specific workload, excessive I/O and cache usage for small operation may be too big.