Know more about "os thread startup"   'os thread startup' takes significant amount of time in 'create index parallel'. All slaves are allocated one by one in serial. SQL tracing on foreground, there is one 'os thread startup' wait per slave, each wait takes 100ms. --> May need investigation When there are 512 slaves, 'os thread startup' wait take 50 seconds before the slaves start to do any job. Resolution is to set *.parallel_min_servers=512 to pre-allocated 512 slaves per instance duirng instance startup, or to run PCIX twice and ignore the first run