This library exposes hardware performance counters via Node.js; benchmarking libraries could use this to support a wider range of performance analysis, and create more reliable (repeatable) benchmarks for Node.js applications.
This error is typically caused by the default permissions for unprivileged users on newer Linux systems. These permissions are controlled by the perf_event_paranoid sysctl file and can be set like so:
sudo sysctl -w kernel.perf_event_paranoid=1
Hardware counters are generally unavailable in virtualized environments. However the relavent features are enabled on some 'bare metal' AWS instances.
Generally speaking to verify that performance monitoring is enabled check the arch_perfmon CPU flag is present:
grep -c "arch_perfmon" /proc/cpuinfo
A result > 0 should indicate that performance monitoring is enabled.
System noise caused by OS jitter and other confounding factors such as CPU frequency scaling and temperature can effect benchmark reproducibility. 1
For this reason a count of CPU cycles is, in general, a more reliable statistic than a timer as a means of measuring and comparing the amount of 'work' a CPU undertook to execute some software.
Many statistical methods have been developed to try to overcome problems of benchmarking, but these are beyond the scope of this package.
A simple way to reduce the effect of several common sources of system noise, however, is to use the Linux cset tool. This tool can confine a program to a single pysical CPU core.
For example:
# create a shield to run the analysis on a physical # CPU (2 logical processors)
sudo cset shield --cpu=2-3 --kthread=on
# run the analysis
cset shield --exec node -- \
./node_modules/.bin/teenytest 'tests/*.test.js'
# reset the shield
cset shield --reset