Software Tuning, Performance Optimization & Platform Monitoring
Discussion regarding monitoring and software tuning methodologies, Performance Monitoring Unit (PMU) of Intel microprocessors, and platform updating.

RAPL misunderstanding / misinformation

Jarvis_S_
Beginner
1,500 Views

Hi All,

Few things still unclear to me regarding RAPL and it seems like the information available online is still unclear,
So I'd be happy if someone could shed some light in regards:

1) RAPL - is that a *real!* measurment, or an estimation taking in account that the chip internals are fixed (in terms of the gates and the conductors), and since intel have this internal info, they can simply estimate and calculate?

1.1) the DRAM measurement : is that estimation as well? How can the CPU / chip, which reports the RAPL, know how much power the DRAM took? (different DRAM vendors. stability across the bus, each module vendor will have a somewhat different cost (above the minimal cost) due to engineering difference).    are these all estimations as well?

2) PSys on skylake - sources on the net says its the SoC,  but on desktops / severs, is it means the entire motherboard consumption? (i.e : all the PCIe slots, the DRAMs, the CPU package and the external auxiliary power that the CPU have?   or - only the motherboard (only the ATX connector) ? )

3) Haswell - is DRAM supported only on XEONs or also on desktops? 

Thanks in advance!! 

 

0 Kudos
6 Replies
Thomas_G_4
New Contributor II
1,500 Views

I can only answer the question 3 with absolute certainty. The Haswell desktop supports the DRAM RAPL domain:

$ likwid-powermeter 
--------------------------------------------------------------------------------
CPU name:	Intel(R) Core(TM) i7-4770 CPU @ 3.40GHz
CPU type:	Intel Core Haswell processor
CPU clock:	3.39 GHz
--------------------------------------------------------------------------------
--------------------------------------------------------------------------------
Runtime: 2.0009 s
Measure for socket 0 on CPU 0
Domain PKG:
Energy consumed: 10.1117 Joules
Power consumed: 5.05357 Watt
Domain PP0:
Energy consumed: 0.499939 Joules
Power consumed: 0.249857 Watt
Domain PP1:
Energy consumed: 0.0411987 Joules
Power consumed: 0.0205901 Watt
Domain DRAM:
Energy consumed: 3.01324 Joules
Power consumed: 1.50594 Watt
--------------------------------------------------------------------------------

A difference between Haswell desktop and Xeons is that they use different energy units for the DRAM RAPL domain

0 Kudos
Thomas_W_Intel
Employee
1,500 Views

The answer to (1) and (1.1) is that power is mostly measured using voltage regulator current monitoring. The execption is Sandy Bridge/Ivy Bridge, where CPU RAPL uses an estimate.

Details can be found in "Energy Efficient Servers - Blueprints for Data Center Optimization" page 62 and 65.

0 Kudos
Thomas_G_4
New Contributor II
1,500 Views

Maybe an answer to (2). The documentation of the PLATFORM domain is not 100% clear but better than nothing:

Total energy consumed by all devices in the platform that receive
power from integrated power delivery mechanism, Included
platform devices are processor cores, SOC, memory, add-on or
peripheral devices that get powered directly from the platform
power delivery means. The energy units are specified in the
MSR_RAPL_POWER_UNIT.Enery_Status_Unit.

And more:

This MSR is valid only if both platform vendor hardware implementation and BIOS enablement support it. This MSR will read 0 if not valid.

Found in SDM Vol3 Chapter 35.15 in the bit description field of MSR_PLATFORM_ENERGY_COUNTER. My Intel Skylake on an Epox board always returns zero.

0 Kudos
Jarvis_S_
Beginner
1,500 Views

Thank you all for your fast and professional responses. 

Thomas Willhalm, regarding what you wrote about 1 and 1.1:

1) so package is the entire of the package even though it is fed by several sources (afaik..?!???  I mean, the motherboard probably feeds some, and the "p8" auxiliary 12v connector)?

1.1) the dram as well is an actual measure rather than an estimate, but how can the Dram be measured from within the cpu (measured, not estimated), if the cpu does not feed the memory power but rather the motherboard itself...?? It will need some sort of physical sensing of the sourcing current to all of the dimm modules on the motherboard... Unless i'm missing something here.  That is why I originally asked about Psys which might include mem as part of the total

I'm on haswell desktop i7, if that somehow relevant.  

Thanks in advance!

0 Kudos
Jarvis_S_
Beginner
1,500 Views

Thomas Willhalm (Intel) wrote:

The answer to (1) and (1.1) is that power is mostly measured using voltage regulator current monitoring. The execption is Sandy Bridge/Ivy Bridge, where CPU RAPL uses an estimate.

Details can be found in "Energy Efficient Servers - Blueprints for Data Center Optimization" page 62 and 65.

 

Thomas, in page 216 of that same book, table 7-1 states that : 

DRAM energy Fixed free-running counter : DRAM energy provides an estimate of the energy used by the DDR3/4 memory devices in the system. Note that support for DRAM RAPL does require some platform enabling, and not all systems support DRAM RAPL.

Can you please clarify this confusion ?  I can't find any critical info in regards.

0 Kudos
Thomas_W_Intel
Employee
1,500 Views

I'm sorry but I may have completely misled you. The book is about servers and it may be all different on a desktop processor.

However, in the end, you need to decide for yourself if the numbers are precise enough for you - regardless of whether they are measured or a good estimate. An DIMM interposer for power measurements should be more precise, but I have no idea if you would be willing to invest for that.

0 Kudos
Reply