3 ms·

> Applications that have high syscall rates include proxies, databases, and others that do lots of tiny I/O. Also microbenchmarks, which often stress-test the s

by mrep 9y ago

> Applications that have high syscall rates include proxies, databases, and others that do lots of tiny I/O. Also microbenchmarks, which often stress-test the system, will suffer the largest losses. My team's RDS instances got hit hard with a 40% increase in CPU usage: https://imgur.com/a/khGxU https://imgur.com/a/khGxU

Xorlev 9y ago
Clarification of the parent: A 40% increase over baseline (60% util -> 84% util), not an absolute 40% CPU util increase. Still terrible though. :(
vanderZwan 9y ago
> Also microbenchmarks, which often stress-test the system, will suffer the largest losses. Tangent: this only seems like more proof that we should not rely on them as developers, no?
- pdpi 9y ago
  It’s proof they serve a specific purpose. If the micro benchmark is especially affected by the effect you’re trying to measure, it provides an upper bound of sorts for the effect’s magnitude. This informs the rest of the analysis.