PYTHON · NUMPY

The counter that cannot miss

A part-time power sensor misreads a periodic workload. You increase its sampling rate five times over. What happens to the error?

Show that the cure for a part-time power sensor is not a faster sensor but a different instrument, by reading an integrating energy counter once and beating a sampler that read five hundred times.

01Challenge

The previous lesson made one workload report 400 W and 100 W by moving nothing but the sampling phase. The obvious fix is to sample faster. Try it, measure what the rate actually buys, and then measure the instrument that does not have the problem at all.

Predict firstA part-time power sensor misreads a periodic workload. You increase its sampling rate five times over. What happens to the error?

02Model

A power sensor answers the question how many watts are flowing right now, and an average of those answers is only as good as the moments it happened to ask. An energy counter answers a different question: how many joules have passed since I started. That second number is a running total the hardware keeps for itself, so the interval between two reads is not a gap in the record, it is the content of the record. Subtract two reads and you have the exact energy of everything that happened in between, including the parts nobody was watching. This is what RAPL is on Intel and AMD parts, and it is why a tool built on it reports the energy of a period rather than a sample from inside it. Alumet, from the Grenoble computer science laboratory with Bull, is built on exactly that and can also reconfigure its acquisition frequency while running, up to ten thousand hertz. The interesting result is that the frequency is not what saves it. You are going to build the counter, run both instruments up a ladder of rates, and find the rate at which the sampler finally becomes correct.

03Guided practice

1. The accumulator is given. energy_counter(trace) returns joules since the start, one entry per step boundary, so acc[b] - acc[a] is the exact energy between steps a and b. 2. Read it, do not sample it. Walk the read points at read_every and add up acc[t] - acc[t - read_every]. Those differences are what a real counter hands you. 3. Divide by time, not by count. The denominator is the elapsed seconds the reads cover, which is the number of steps covered times DT, not the number of reads. 4. Then read the ladder. Look at the 0.37 duty rows before the 0.50 ones. A duty cycle that does not divide the sampling grid is the honest case, and it is where the rate stops helping.

04Feedback

PASS WHEN You read an energy counter once and got the exact mean power, while a sampler taking five hundred readings of the same trace was still wrong.

If your run fails, check:

Divide by the elapsed time, not by the number of reads. If your answer scales with read_every, that is the error.
Use the differences of the accumulator, not the accumulator values themselves. A running total is not a power reading.
Define counter_mean_power(trace, read_every) returning one number in watts.

05Retrieve & space

At duty 0.37 the sampler reads 220 W with both 100 and 500 reads. Why did five times the rate buy nothing?
The sampler is exact at 2000 reads. What has it stopped being at that rate, and what did that exactness cost?
Your workload has no period at all, just random bursts. Which of the two instruments is still exact, and why does the answer not depend on the workload?

06Mastery & project

You can say which of the two questions an energy figure answers, how many joules passed or how many watts were flowing at some instants, and therefore whether a sampling rate is a relevant thing to ask about it at all.

Find out which one your own machine gives you. Read the instrument's documentation for the words energy, accumulated or since reset, as against power, instantaneous or current. If it is a counter, state the read interval and stop worrying about phase. If it is a sampler, state its duty cycle beside every figure you publish. The distinction is a grade of evidence and not a detail of technique: a figure read from a counter has been measured, and a figure integrated from samples has been derived, and the two do not belong in the same column of the same table.

GO FURTHER · THE LIVE PLAYGROUND

Alumet ↗ · The Grenoble computer science laboratory with Bull: an energy monitoring agent on RAPL, reconfigurable while running, up to 10 kHz. It reports the energy of the interval between reads, which is the instrument this lesson builds.
The meter that sleeps ↗ · The defect this lesson answers: one workload reporting 400 W and 100 W against a true 250 W by phase alone.
PAI-350 · Reading an Energy Claim ↗ · The same question asked of a figure somebody else published.

← The meter that sleeps Back to course →

PYTHON · NUMPY

The counter that cannot miss

One duty-cycled workload, two instruments, five sampling rates, and one number that does not move.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you