29 Aug
2014
29 Aug
'14
5:37 p.m.
On Thu, 28 Aug 2014 23:40:01 +0000
Kaza Kore <dj_kaza(a)hotmail.com> wrote:
So can anybody point to any conclusive evidence that
i-series
processors benefit from having HT disabled on a Linux based DAW?
Preferably benchmarks on a system installed with HT Enabled and
Disabled using a recent kernel and system.
Afaik there are 2 points to be made.
1. SMT (at least as implemented on iX cpus) break the entire concept
of realtime. You do get an estimated increase in CPU throughput of
about 25%, at the cost of stealing CPU time from your SCHED_FIFO/RR
thread. What it in effect means is that a SCHED_NORMAL thread can run
on the sibling CPU stealing CPU time from your SCHED_FIFO/RR thread.
2. I suspect that the CPU instruction cache is too small. and that
SMT can also cause cache depletion, which would explain some xruns I
have seen while torture testing the system doing real audio and
running hackbench at the same time.
Speedstep was already covered by someone else.
CPU management is a thorny question. That setting is probably
related to System Management Interrupts (SMIs), which can be very bad
indeed. Normally used for fan control and other functions by the
BIOS. The really bad thing is that they block execution and there is
nothing the kernel can do about it. if it's only for a few usecs,
it's probably insignificant and not a problem, but if it's
milliseconds, then you will get xruns...
There is a program called hwlatdetect in the rt-tests package that can
be used to determine how big the problem is. It consists of a kernel
module and a python script to start it and to report the results back
to the user. What is it does is basically to stop all kernel execution
and then to loop reading TSC timestamps. If it finds breaks in the
data stream they will have been caused by SMIs.
--
Joakim