On 30 Mar 2014 at 14:00, Bjorn Reese wrote:
Out of curiosity, how many CPU cycles per op in your ASIO test case? AFIO seems to need ~9,000 CPU cycles per op processed, half of which is spent spinning on CAS locks - I would assume that ASIO can knock that down by two thirds?
1000 cycles/op -- measured via io_service::do_run_once().
Just to clarify, my ~9000 cycles/op is for maximum contention i.e. fully loaded with eight threads all fighting it out. Is your 1000 cycles/op for two threads only? Methinks AFIO could do with some minimum latency benchmarks actually ... might as well, I already have build time benchmarks.
It locks/unlocks four times per operation, which accounts for a total of 20% of the CPU time.
I'm actually surprised it's as much as that. I would have thought twice per operation is the minimum possible, but you have to make some hard design choices to get it that low. AFIO "looks funny" partially because it locks exactly twice per op as from the v1.2 engine, unless you have TSX in which case it never locks at all except if more memory from the kernel is needed. Niall -- Currently unemployed and looking for work in Ireland. Work Portfolio: http://careers.stackoverflow.com/nialldouglas/