Unless evil gremlins abound (which would also make logical analysis pointless) it seems inconceivable that the print can in itself cause
that much variation in performance. I would be inclined to look elsewhere, such as network issues. If it weren't for your timer evidence I'd have suspected multiple queueing for that db lock and even now it sort-of galls me to have to eliminate that - it seems oh-so credible as a suspect cause, given the "arithmetic" variation in performance measurements you report that suggest a per user cumulative delay - way tempting!