7 Comments

It would be interesting to add power figures - a major selling point for opcache is that it allows to power down the decoders and fetch and save a lot of energy.

So even though the FP workloads dont benefit from opcache directly, they may benefit greatly from the decteased power consumption of front end.

Expand full comment

I know :/ But I don't have power measuring hardware, and research papers indicate AMD's CPU core/package power figures may be modeled, not measured. For something that likely makes milliwatts of difference, I'm not sure I can figure anything out from those package/core power counters.

Expand full comment

Can you do similar tests on newer Intel CPUs? Do they let you disable the opcache?

Also, it would be fascinating to do this sort of thing across a wide collection of cpu generations. Going back as far as possible :D

Expand full comment

In tight compute loops is there any reduction in power consumption? Hot spots can develop in tight loops.

Expand full comment

That would require access to some very accurate and fast measuring devices, which as far as I know they don't have. The biggest "efficiency gain" seems to come from making computation faster. Ergo using slightly less energy (?) for much shorter time. Though I'm not even sure if this lowers energy use per se, since it is active silicon.

Expand full comment

It'd be lovely to have such low-level metrics exposed in such way they could end up in Grafana dashboards for a server farm (or home lab). Do you have any ideas of how that could be done (assuming someone didn't already do it)?

Expand full comment

Not sure if anyone else already did it. You'd have to read MSRs or use perf's API, then somehow push those values to Grafana or whatever application runs a dashboard

Expand full comment