[LAU] CPUs for low latency

[Len Ovens]

Just looking through some of the results on:

https://www.osadl.org/Hardware-overview.qa-farm-hardware.0.html

Some of the higher speed many core CPUs have the worst latency:
https://www.osadl.org/Latency-plot-of-system-in-rack-1-slot.qa-latencyplot-r1s1.0.html?shadow=0
https://www.osadl.org/Latency-plot-of-system-in-rack-0-slot.qa-latencyplot-r0s7.0.html?shadow=0
(These will cause xruns at 16/2 at 48000 sample/sec)

Some slow cpus do much better:
https://www.osadl.org/Latency-plot-of-system-in-rack-3-slot.qa-latencyplot-r3s2.0.html?shadow=0
https://www.osadl.org/Latency-plot-of-system-in-rack-1-slot.qa-latencyplot-r1s4.0.html?shadow=0

Two very similar CPUs can vary:
https://www.osadl.org/Latency-plot-of-system-in-rack-c-slot.qa-latencyplot-rcs6.0.html?shadow=0
https://www.osadl.org/Latency-plot-of-system-in-rack-0-slot.qa-latencyplot-r0s2.0.html?shadow=0
(the second is a faster but older one)

Hyperthreading does not effect latency?:
https://www.osadl.org/Latency-plot-of-system-in-rack-0-slot.qa-latencyplot-r0s4.0.html?shadow=0
(at least this plot seems to say that) The state of HT and what effects 
how much it hits latency has changed. In order for HT to be worth while, 
any one thread needs to keep the core for a certain number of 
instructions. As the CPU speed goes up the same number of instructions 
gets done faster. I also think that the overhead for context switching has 
gone down. 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.

AMD seems to have something similar enough to hyperthreading to be wary 
of. A number of the AMD CPUs shows twice as many threads as cores. I know 
that they do double caching, I don't know if this is what the Linux kernel 
is seeing as shadow cpus or not.

Looking through some of the AMD documentation (which is very sparse BTW) 
It appears to say that some of the boost and processor speed changes are 
not controllable by bios.

On any of the Intel CPUs, both boost and HT can be turned off by any 
reasonable bios. With AMD I am sure that the kernel at least can be told 
to ignore shadow CPUs. I would like to know more about speed scaling and 
how much influence the the OS has on it. Another thing with CPUs that 
seems to becoming common is powering off parts of the CPU that are idle. 
The poweron time is very short, but context change time would still be 
there. There have been some experiments that show just keeping the cpu 
busy with a do-nothing-script can give lower latency times.

Anyway, the latest, greatest MB and CPU is not the best bet for low 
latency audio. Advertised "performance" is based on throughput. Any of the 
benchmarks I have seen on the manufacture's sites are about throughput. 
Latency is not mentioned or measured. The few times it has been mentioned 
has been in regard to video where it is called "low latency" which 
probably in the audio world is noticable delay. (in most computer 
engineer's minds low latency is 30ms)

The idea of what latency is good enough for audio work has changed with 
the change from mostly PCI cards to USB. Where 64/2 was the normal 
maximum, 128/2 or even 256/2 have become common... and seem to be accepted 
as "OK". (there are also USB cards that will run at 32/2 or better) Maybe 
there are other reasons for the change in thought on this too.

Note: this was a very lite look, not in depth at all. The profiles do not 
match the tests always. I see that in the profiles the governor is set to 
OnDemand, but the test plots all say "lowest P state: performance".

--
Len Ovens
www.ovenwerks.net