I have an Asrock C2750D4I that I use for a 8x3TB SATA ZFS NAS and recently bought a LSI 9217-16i SAS HBA so I could stop using the Marvell controllers on the motherboard.
I wiped everything and before I put in the LSI controller and I ran a few benchmarks with a fresh 11.0-p1 install (running of an SSD on the SoC Sata III) and tested the 8x3TB drives in a flat, mirror and raidz2 layout (4k with gnop, raidz2).
I then moved all 8 drives over to the LSI HBA, imported the pool and reran the tests (phoronix pts/disk). The last config was raidz2 and I was surprised to find that the tests ran slower on the LSI card. On a hunch I wiped the pool and recreated the raidz2 pool with the same settings and reran the tests. The tests ran better (slightly faster than any of the zfs layouts with the onboard controllers). This is what I originally expected.
My question is, why did the imported pool run slower? The drive geometry doesn't change regardless of the controller so I'm a little curious what the difference is importing a pool from a different controller instead of creating it on the current controller. For what it's worth, the original pool was spread across three controllers. Four drives were on a Marvell 88SE9230, two on a Marvell 88SE9172 and two on the Sata II (3Gbps) built into the C2750 SoC.
Anyone have any insight/thoughts?
I wiped everything and before I put in the LSI controller and I ran a few benchmarks with a fresh 11.0-p1 install (running of an SSD on the SoC Sata III) and tested the 8x3TB drives in a flat, mirror and raidz2 layout (4k with gnop, raidz2).
I then moved all 8 drives over to the LSI HBA, imported the pool and reran the tests (phoronix pts/disk). The last config was raidz2 and I was surprised to find that the tests ran slower on the LSI card. On a hunch I wiped the pool and recreated the raidz2 pool with the same settings and reran the tests. The tests ran better (slightly faster than any of the zfs layouts with the onboard controllers). This is what I originally expected.
My question is, why did the imported pool run slower? The drive geometry doesn't change regardless of the controller so I'm a little curious what the difference is importing a pool from a different controller instead of creating it on the current controller. For what it's worth, the original pool was spread across three controllers. Four drives were on a Marvell 88SE9230, two on a Marvell 88SE9172 and two on the Sata II (3Gbps) built into the C2750 SoC.
Anyone have any insight/thoughts?