Re: Power-Mgt (Was: Re: cvs commit: src/sys/i386/cpufreq est.c )

On Mon, 17 Mar 2008, Poul-Henning Kamp wrote:

In message <20080317141717.U3253@xxxxxxxxxxxxxxxxx>, Robert Watson writes:

If cpufreq is going to be enabled by default, should we be enabling powerd by
default [...]

[Moved to arch@]

In general, I think we must make power-aware computing our "next
SMPng project", not in the sense of delaying the next major release
five years, but in the sense that power consumption should permerate
our thinking about the operating system from now on.

Overall, I think that means that we should:

* Enable performance neutral power savings on servers
- spin down unused disks. (geom/drivers)
- use only as many CPU cores as necessary (scheduler)

This is an interesting notion which I have tried to leave room for in the current scheduler design. One thing which I have considered in the past is a policy of best power vs best performance.

For example, consider a multi-socket system with multi-core parts. With two, unrelated, runnable threads, you'll get the best perf by putting them on different sockets. Then they'll have the most cache and memory bandwidth available to them. You'd be able to spin down a socket if you put them on adjacent cores on the same socket. It's not clear that this would be a power savings however, what if each thread now runs at half the speed? Is that more power efficient than running two cores half the time?

And what about the barcelona, which can power down individual cores and even individual parts of cores? And in this point to point bus topology you always need to have the dram controler and HT link on anyway.

One further complication is of course that cpus can idle in different states. So someone really is going to have to explore the tradeoff between core speed, number of cores, power and performance.

I think the answer to which scheduling algorithm is most power efficient is really going to come down to the cpu architecture and type of workload. This is why I have been reluctant to implement anything yet. I suspect that getting things done the fastest is going to be a good first approximation of using the least power in this regard.



- light cpu-throttling.
- downgrading 1GB to 100MB ether when idle.

* Aim to meet or execeed energystar 4.0/5.0[1] on desktops and
plugged laptops.
- Pretty much as above, but with specific targets.
- http://www.energystar.gov/index.cfm?c=revisions.computer_spec

* Be as battery-frugal as possible on battery driven laptops.
- Any trick in and off the book.

I think these are all good goals. You should also throw in there the tickless time keeping that linux has done. We've talked about it for ages too but never gotten there. It's a shame that we keep playing catchup in these areas.

Thanks,
Jeff


--
Poul-Henning Kamp | UNIX since Zilog Zeus 3.20
phk@xxxxxxxxxxx | TCP/IP since RFC 956
FreeBSD committer | BSD since 4.3-tahoe
Never attribute to malice what can adequately be explained by incompetence.
_______________________________________________
freebsd-arch@xxxxxxxxxxx mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-arch
To unsubscribe, send any mail to "freebsd-arch-unsubscribe@xxxxxxxxxxx"

_______________________________________________
freebsd-arch@xxxxxxxxxxx mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-arch
To unsubscribe, send any mail to "freebsd-arch-unsubscribe@xxxxxxxxxxx"

Relevant Pages

  • Re: Performance improvement using X5355 over 5080
    ... It runs Windows Vista x64 and scores 5.9 in all areas of the Windows Experience Index except for memory where the latency of FB-DIMMS results in a score of only 4.7. ... I guess it's hard to say what magnitude of performance increase would result from a 33% increase in FSB but it obviously wouldn't be a corresponding 33% increase. ... real cores or 8 HT cores and is replacing in with 8 real cores). ... Also, the 5355 use quite a bit of power, they don't use quite as much ...
    (comp.sys.intel)
  • Re: Performance improvement using X5355 over 5080
    ... X5355 2.66GHz Quad-Core CPUs? ... real cores or 8 HT cores and is replacing in with 8 real cores). ... For an parallelizable workload that doesn't stress memory too hard I'd ... Also, the 5355 use quite a bit of power, they don't use quite as much ...
    (comp.sys.intel)
  • Re: Rumours of (CPU) Wars
    ... >> the die that the aggregate chip performance will be faster. ... referred to the aggregate performance of all the cores on the chip was quite ... improve power consumpation at all) emphasizes the lesson that within a given ... IBM will simply lower the price accordingly. ...
    (comp.os.vms)
  • Re: fixed point math library?
    ... The compiler can only optimize if the promotions don't care, ... Power consumption and die size are important factors to me. ... ARM cores in a square mm), because they've been optimized for speed, ... bit add or multiply in 1 compact instruction, ...
    (comp.arch.embedded)
  • Re: VMS backup: competition from Apple
    ... What will a desktop PC do with 32 cores? ... absolutely nothing to do with competing with Alpha. ... Sun did multi-core because it saved power, ...
    (comp.os.vms)