On 08/19/12 3:16 PM, Smithies, Russell wrote:
For a high-performance system (64-cores, 512GB RAM, 5TB local disk, 110TB NFS-mounted storage) is there any advantage of dropping lvm and mounting partitions directly? We're not planning on changing partition sizes, but if we did we'd probably do a full rebuild. Has anyone done performance testing to show that lvm isn't crippling I/O?
running an extensive series of pgbench (postgresql benchmark similar to tps-c) with and without lvm and using ext4 vs xfs, on a very large raid10 built from 16 drives, the performance differences were pretty much non-existant, down in the noise, less than 1%. this was on a 12 core (24 thread) 3Ghz 48GB ram system, using a LSI logic megasas2 raid card with 512MB battery backed writeback cache. the IO in these tests were 99% random write, and the tests were run with all 4 combinations of LVM or not and EX4 vs XFS and with different client connection counts (typically up around 100 client connections was where we got the peak transaction throughput
the entire raid was dedicated to the database tests, the OS was running on a seperate raid1 which had pretty much zero IO during the benchmark sequences.
now, this isn't saying that you might not get a totally different result under different test conditions.