Commit7db0dc0

committed

pgbench: Comment on thread timing hazards.

Reviewed by Fabien COELHO.

1 parentbf26179 commit7db0dc0Copy full SHA for 7db0dc0

File tree

+10

-1

lines changed

+10

-1

lines changed

Lines changed: 10 additions & 1 deletion

Original file line number	Diff line number	Diff line change
`@@ -2896,7 +2896,16 @@ main(int argc, char **argv)`
`2896`	`2896`	`}`
`2897`	`2897`	`disconnect_all(state,nclients);`
`2898`	`2898`
`2899`		`-/* get end time */`
	`2899`	`+/*`
	`2900`	`+ * XXX We compute results as though every client of every thread started`
	`2901`	`+ * and finished at the same time. That model can diverge noticeably from`
	`2902`	`+ * reality for a short benchmark run involving relatively many threads.`
	`2903`	`+ * The first thread may process notably many transactions before the last`
	`2904`	`+ * thread begins. Improving the model alone would bring limited benefit,`
	`2905`	`+ * because performance during those periods of partial thread count can`
	`2906`	`+ * easily exceed steady state performance. This is one of the many ways`
	`2907`	`+ * short runs convey deceptive performance figures.`
	`2908`	`+ */`
`2900`	`2909`	`INSTR_TIME_SET_CURRENT(total_time);`
`2901`	`2910`	`INSTR_TIME_SUBTRACT(total_time,start_time);`
`2902`	`2911`	`printResults(ttype,total_xacts,nclients,threads,nthreads,`

Comments

(0)