Overhead of Measurement Providers

The Green Metrics Tools measurement providers run on the same system as the software to be measured.

This allows for easiness of testing, but poses the risk of skewing the measured results.

The reporters are designed to have negligible impact on their own.

Results for one measurement provider #

Tests done with powerstat

• All tests executed remotely via SSH
• Values format is: Average - Std.Dev
• powerstat call was “sudo powerstat -R -c -D 1 60”.
• 60 samples in 1 s interval

Average on idle

• PKG Watts: 2.41 - 0.06
• DRAM Watts: 0.46 - 0.00
• CPU Idle %: 99.5 - 0.4

measurement provider 10ms to terminal / STDOUT

• PKG Watts: 2.85 - 0.13
• DRAM Watts: 0.5 - 0.01
• CPU Idle %: 99.4 - 0.5

measurement provider 1ms to terminal / STDOUT

• PKG Watts: 4.67 - 0.12
• DRAM Watts: 0.55 - 0.04
• CPU Idle %: 97.7 - 0.7

measurement provider 10ms to /dev/null via >

• PKG Watts: 2.69 - 0.39
• DRAM Watts: 0.48 - 0.02
• CPU Idle %: 99.3 - 1.2

measurement provider 1ms to /dev/null via >

• PKG Watts: 3.42 - 0.40
• DRAM Watts: 0.49 - 0.02
• CPU Idle %: 98.4 - 0.8

measurement provider 10ms to file in C (fprintf)

• PKG Watts: 2.65 - 0.07
• DRAM Watts: 0.47 - 0.00
• CPU Idle %: 99.6 - 0.4

measurement provider 1ms to file in C (fprintf)

• PKG Watts: 3.32 - 0.30
• DRAM Watts: 0.49 - 0.02
• CPU Idle %: 98.1 - 0.8

measurement provider 10ms to file via >

• PKG Watts: 2.86 - 0.06
• DRAM Watts: 0.48 - 0
• CPU Idle %: 99.5 - 0.3

measurement provider 1ms to file via >

• PKG Watts: 3.49 - 0.35
• DRAM Watts: 0.5 - 0.03
• CPU Idle %: 98.3 - 1.3

Summary #

The most costly output is via STDOUT, which is only done if you run the measurement provider manually for testing.

In the Green Metrics Tool output is always redirected via POSIX redirection to a file. This mode has the lowest impact, however not distinguishable from writing to file directly or to /dev/null

Since we recommend running the reporters at 100ms for normal analysis we conclude that we have a lot of headroom by just running one reporter.

Running multiple reporters / Full Green Metrics Tool #

We have done measurements on two machines with a standard setup:

Quanta Leopard-DDR3 48-Core #

• Normal Idle without GMT: 60 W
• Full load: 300 W

then we activated the following reporters:

• psu.energy.ac.ipmi.machine.provider.PsuEnergyAcIpmiMachineProvider: 100 ms resolution
• network.io.cgroup.container.provider.NetworkIoCgroupContainerProvider: 100 ms resolution
• cpu.energy.RAPL.MSR.component.provider.CpuEnergyRaplMsrComponentProvider: 100 ms resolution
• cpu.utilization.procfs.system.provider.CpuUtilizationProcfsSystemProvider: 100 ms resolution
• memory.total.cgroup.container.provider.MemoryTotalCgroupContainerProvider: 100 ms resolution
• memory.energy.RAPL.MSR.component.provider.MemoryEnergyRaplMsrComponentProvider: 100 ms resolution
• cpu.utilization.cgroup.container.provider.CpuUtilizationCgroupContainerProvider: 100 ms resolution

The power draw is now 63-66 W

This means the GMT reporters have around a 1-2 % overhead regarding the full power draw.

Fujitsu TX1330 M2 #

• Normal Idle without GMT: 15.8 W
• Full load: 50 W

then we activated the following reporters

• network.io.cgroup.container.provider.NetworkIoCgroupContainerProvider: 100 ms resolution
• cpu.energy.RAPL.MSR.component.provider.CpuEnergyRaplMsrComponentProvider: 100 ms resolution
• cpu.utilization.procfs.system.provider.CpuUtilizationProcfsSystemProvider: 100 ms resolution
• memory.total.cgroup.container.provider.MemoryTotalCgroupContainerProvider: 100 ms resolution
• psu.energy.ac.powerspy2.machine.provider.PsuEnergyAcPowerspy2MachineProvider: 250 ms resolution
• memory.energy.RAPL.MSR.component.provider.MemoryEnergyRaplMsrComponentProvider: 100 ms resolution
• cpu.utilization.cgroup.container.provider.CpuUtilizationCgroupContainerProvider: 100 ms resolution

The power draw is now 16.3 W

This means the GMT reporters have around a 1 % overhead regarding the full power draw.