Syed Amer Gilani wrote: > I am currently in the process of writing a report about how well the > realtime results are with xenomai for a project. For this i am of > course using latency and cyclictest. But the results are somewhat odd. > With Load from dd and external ping -f, latency returns results i am > expecting: http://systec.de/dump/latency-load.png > But without Load there is a odd "hole" from 3us to 5us: > http://systec.de/dump/latency-noload.png > Can this somehow be explained?
Maybe a cache granularity effect: either the caches are hot and can serve most requests of the benchmark directly so that you get latencies <3 us, or parts of them have been overwritten, and it takes typically >5 us to replace the required content. > cyclictest also has this hole: http://systec.de/dump/cyclictest.png > But i am not sure if i completely understand what cyclictest does. Is > it the same as latency just for the posix skin? Yes and no. latency can benchmark timed user space, kernel space, and IRQ handler scenarios, but only for one event per program instance. cyclictest can also run multiple events (tasks), but only in user space. latency is using native+rtdm skin, cyclictest the posix skin. So the best comparison between both tools is running a single task in user space at the same interval with each of them (and let them run roughly the same time...). Jan
signature.asc
Description: OpenPGP digital signature
_______________________________________________ Xenomai-help mailing list [email protected] https://mail.gna.org/listinfo/xenomai-help
