Internal performance profiling indicates that maglev_lb.cc consumes some of the most CPU cycles for jobs running the Envoy binary.

Taking a closer look at maglev_lb.cc, we can see that there is an expensive modulo operation for every iteration over the maglev table. Some possible explanations for this cost is

• the default maglev table size being 65537 setting a large minimum number of iterations
• changing backends can trigger table rebuilds often

One possible way to improve the efficiency of building the maglev table is to avoid the use of modulo.

Benchmarks indicate a possible 15% improvement in efficiency:

Baseline

Avoid use of modulo