I guess we should have both.

Real-world user perceived time is wall-time not CPU time, so we should definitely keep that in the spirit of real-world benchmarks.

yeah, having both is fine.

As discussed with fmeawad a while ago: using cpu time doesn't help much with reducing noise.

hum, now that we actually have cpu time, the noise in my graphs is down to an acceptable level.

before, i got a bimodal distribution, depending on whether the v8 execution or something else that happens during load got descheduled.

oh ok :) then let's keep this open.

well, with tbmv2 it's actually fixed :)