Presentation is loading. Please wait.

Presentation is loading. Please wait.

Matthew Garrett <mjg@redhat.com> Going green with Linux Matthew Garrett <mjg@redhat.com>

Similar presentations


Presentation on theme: "Matthew Garrett <mjg@redhat.com> Going green with Linux Matthew Garrett <mjg@redhat.com>"— Presentation transcript:

1 Matthew Garrett <mjg@redhat.com>
Going green with Linux Matthew Garrett

2 Data centres are expensive
~3% of total US electrical consumption in 2010 ~15GW ~15 nuclear power stations

3 Fixing infrastructure
On average, 1W of overhead for 1W of machine More efficient cooling systems More efficient power supplies

4 Things are getting better!
0.2W of overhead per W of machine

5 Flying car future Use waste heat to generate power

6 Linux doesn't solve this
(Sorry)

7 Things need to get a lot better...
High end hardware takes a lot of power 48 core system = ~350W at idle (with power management...)

8 What can we do in software?
CPU power management dominates right now Modern x86 idles at ~0W Full power = ~20W per core (48 core system = 960W from CPUs alone!)

9 Gettings CPUs to do less is key
An idle CPU is a cheap CPU Full package sleep means idle memory ~5W per stick of DDR3 at load, ~0.1W in self refresh Optimising CPU and memory use patterns makes a huge difference

10 Idle CPU cores = idle CPU package
RAM

11 One active core = active package
RAM

12 Keeping idle is key

13 Keeping idle is key

14 Keeping idle is key

15 Huge efforts to reduce wakeups
IPMI reworked Automounter moved to event driven model Application-level tuning Debug tools (powertop, systemtap)

16 What are the numbers? At fully idle, Linux ~3% better than Windows
Things start getting interesting under load...

17 (Illustrative results)

18 (Illustrative results)

19 Take home message From 10%-50% load, Linux significantly more power efficient Margin decreases up to 90%, equality reached at 100%

20 Tweaking Linux behaviour
/sys/devices/system/cpu/sched_mc_power_savings

21 RAM RAM

22 One task RAM RAM

23 Two tasks (sched_mc_power_savings = 0)
RAM RAM

24 Two tasks (sched_mc_power_savings = 1)
RAM RAM

25 Two tasks (sched_mc_power_savings = 2)
Process migrated RAM RAM

26 Future trends Smarter scheduling More runtime memory power management
Increasing virtualisation as a power management strategy Power management of smaller system components

27 Scheduling NUMA-aware migration
Consolidate active pages in smaller nodes Thermal-aware scheduling If a CPU is limited because of heat, don't use it

28 Virtualisation Run virtualised rather than on real hardware
Consolidate guests on a smaller number of hosts Bring up hosts and migrate guests as demand increases (requires rapid guest migration)

29 Load balancing Put backend machines to sleep
Wake when demand increases Reduced resume latency

30 What are the tradeoffs? Focus has been on “free” power management
Is it really true?

31 Well... Enterprise customers have enterprise demands
Sometimes that performance hit really does matter

32 Latency can be a pain C-states inherently create latency
Some people really don't want latency There's an obvious conflict here

33 pm_qos Simple solution is to let applications indicate their requirement pm_qos interface provides that

34 Performance/power tradeoff
Not always the direction you'd think Intel's turbo mode depends on good idle PM (Benchmarks in the snow)

35 Turning more off Marginal but measurable gains from caring about unused hardware Lots of machines have USB ports Many don't have anything in them...

36 Genuinely user transparent
USB SD Firewire Audio?

37 Focus on power management
Reducing power consumption reduces costs

38 Focus on power management
Reducing power consumption reduces costs Reducing power consumption is good PR

39 Focus on power management
Reducing power consumption reduces costs Reducing power consumption is good PR Reducing power consumption is common sense

40 Questions?


Download ppt "Matthew Garrett <mjg@redhat.com> Going green with Linux Matthew Garrett <mjg@redhat.com>"

Similar presentations


Ads by Google