On 12/20/2015 06:20 PM, Len Ovens wrote:
[..]
Note that these plots are the best-case. While cyclictest can load the
CPU, it does not perform any FPU computation as DSP would.
Hyperthreading does not effect latency?:
https://www.osadl.org/Latency-plot-of-system-in-rack-0-slot.qa-latencyplot-…
(at least this plot seems to say that)
It's very application specific, HT can reduce latency for some
workloads: mainly application with many branches that cause instruction
cache misses.
Digital Signal Processing however requires complete attention of the CPU
core and does not usually stall (at least well written DSP optimizes
code loops and branches to take instruction fetch into account). In this
case Hyperthreading leads to increased latency since two DSP threads on
the same CPU compete with each other.
Intel-Research published a paper about this:
https://software.intel.com/en-us/articles/performance-insights-to-intel-hyp…
With my old P4 (single core) HT restricted my AI to
64/2
with the ocasional xrun, HT off would allow solid 16/2 performance. My
new i5 does not have HT, so I am not able to compare.
+1 for testing + measuring.
Anyway. In most cases the CPU is not the limiting factor. In the vast
majority of cases some power-saving (bus-freq, southbridge, SMI) or
weird graphics-cards, wifi, etc and/or their drivers are orders of
magnitude more limiting.
2c,
robin