Energy consumption of evolutionary algorithms in JavaScript

Merelo (UGR-ES), García-Valdez (ITT-MX), Castillo (UGR-ES)

Wivace, 2023, Venice

Green computing

Minimize the amount of energy spent in algorithms/applications

No fixed set of rules for "greening"

Algorithm/language specific experimental methodology

JavaScript is here to stay

ECMA standard

And has different interpreters: bun, deno, node.js

Profiling the energy: finding how much energy the system spends

Using RAPL: Intel-defined machine readable machine registers

PKG == CORE + RAM

Using command-line tools that measure single processes

Measure 15 times, take average, measure energy consumption of sleep during that time, substract from measure

Not ideal but roughly accurate

Fitness, mutation and crossover, most important operations

Involves integer arithmetic + list/string manipulation

So far:

Measurements March 2023, published July 2023

bun rocks

pinpoint, best tool for command-line measures

What we'll be doing now

Is there any change with new runtime versions?

Is mutation also a significant operator?

What happens with different hardware architectures?

Is there any change with new runtime versions?

Definitely

How?

It's complicated

Version change: node.js uses less energy

Version change: bun uses less energy... to a point

Version change: deno improves energy and time only for big chromosome sizes

The promissione ducale was renovated for every new doge

Measurements are only valid for a single minor version

Should mutation also be measured ?

Not really

Mutation operator

Let's drop mutation for future tests

What happens with different hardware architectures?

Energy consumption will depend on hardware implementation

What is measured also changes: AMD emulates, Intel does not

Measuring core energy consumption

Measuring RAM energy consumption

Average time (s) and cost (Joules)

VM Size AMD - s AMD - J Intel - s Intel - J
bun 1024 5.074043 141.1920 7.951114 78.30800
bun 2048 7.839581 253.4840 12.520187 117.62600
bun 4096 13.743342 377.8240 24.408003 193.08267
deno 1024 12.607883 400.6260 19.325474 143.78200
deno 2048 9.969686 319.4293 19.350870 70.27667
deno 4096 9.699899 297.8767 20.078916 132.41000
node 1024 11.132792 167.4040 15.823338 100.29200
node 2048 17.748798 460.2467 32.422865 203.13200
node 4096 32.627382 1145.8933 29.285567 228.25267

There's usually a speed/energy efficiency tradeoff

With speed losses overcome by bigger energy savings, usually

Speed and energy savings depending on factors such as size

Conclusions

Energy expenses are minimized using bun and Intel laptops

But measures are needed for every change in platform/interpreters

Any VM consumes less than node.js

Vi ringrazio per la vostra attenzione