raid10 workstation pg_test_fsync

raid10workstation$ pg_test_fsync
2000 operations per test
O_DIRECT supported on this platform for open_datasync and open_sync.

Compare file sync methods using one 8kB write:
(in wal_sync_method preference order, except fdatasync
is Linux's default)
        open_datasync                      81.331 ops/sec
        fdatasync                          82.531 ops/sec
        fsync                                      26.609 ops/sec
        fsync_writethrough                            n/a
        open_sync                          26.430 ops/sec

Compare file sync methods using two 8kB writes:
(in wal_sync_method preference order, except fdatasync
is Linux's default)
        open_datasync                      38.820 ops/sec
        fdatasync                          81.045 ops/sec
        fsync                              26.782 ops/sec
        fsync_writethrough                            n/a
        open_sync                          13.013 ops/sec

Compare open_sync with different write sizes:
(This is designed to compare the cost of writing 16kB
in different write open_sync sizes.)
        16kB open_sync write               26.974 ops/sec
         8kB open_sync writes              13.005 ops/sec
         4kB open_sync writes               6.566 ops/sec
         2kB open_sync writes               2.690 ops/sec

... at this point I got bored waiting of the amazingly, incredibly glacial performance and just posted.


raid10workstation$ mdadm --detail /dev/md1

     Raid Level : raid10
               ...
   Raid Devices : 4
  Total Devices : 4



raid10workstation# hdparm -i /dev/sda

/dev/sda:

 Model=WDC WD1001FALS-00J7B0, FwRev=05.00K05
 ..
 UDMA modes: udma0 udma1 udma2 udma3 udma4 udma5 *udma6
 AdvancedPM=no WriteCache=enabled



Note that while the disk's write cache is shown as enabled, the RAID subsystem uses write barriers to ensure proper flushing. In practice the write cache has little effect.

No LVM is in use. LVM makes performance even worse.

This is software RAID, but hardware RAID in write-through caching mode with no BBU will not be much - if at all - better.