To get a more realistic workload, the next few samples use a much larger scale of bench database (1000) and more clients (64):
$ pgbench -i -s 1000 pgbench $ pgbench -j 4 -c 64 -T 300 pgbench
This gives about 400 TPS on this server. The following snapshot shows one type of problem you can discover from the vmstat data:
$ vmstat 1 procs ----io---- --system--- -----cpu------ r b bi bo in cs us sy id wa st 3 62 4916 34296 11504 23871 8 6 0 85 0 2 64 7132 35828 13247 32406 11 8 0 81 0 4 63 6120 40756 11313 29722 35 7 0 58 0 0 48 3896 13712 6953 19544 28 3 2 66 1 0 12 400 25564 2222 3417 0 1 25 73 0 0 5 44 3700 818 1020 0 0 39 61 0 1 12 64 8364 1388 1773 6 0 44 50 0 1 45 704 7204 2133 3883 23 1 3 73 0 5 60 2912 26148 ...
