Can Anyone Explain Threads In Nuke

[flaviomeibach]

hi everyone,

i realised that nuke is not running on full capacity on my xp / asus p5b / q6600 system with the threads left to 8

the fastest option for me is to set it to 2, which leaves 2 cores completely idle while the other two are running 100% each.

there is no hyperthreading on my system and i disabled all relevant power saving options in the bios. the microsoft hotfix for multiprocessors is also intalled.

i know that nuke does not run quite as good on xp than on linux and i am aware of xp's insufficiant multiprocessor/multicore capabilities

--- so, what is the theory behind the nuke threads, and how does that apply to a multicore system like mine?
--- is there a way i can get nuke to use all cores at full capacity?

is that a problem that applies only to multicore cpus or also to mulitprocessor systems (which could of course also have multiple core cpus)

thanks

[flaviomeibach]

actually, sometimes setting it to just 1 thread is the fastest option.

interestingly, the cpu usage in the taskmanager does not properly reflect the speed that nuke renders with.

with threads set to 1 i sometimes get overall cpu usage of 25% and it renders faster than with 2 threads and cpu usage of 40%

why is that so? what eats my cpu power?

[titane]

I have got the same problem and it seems to me
that Nuke completely ignores any options I set
in the commandline.
so I would be more than happy if anyone could explain
that to me.
btw I am using xp and xp64 on dual opteron systems.

[joojaa]

QUOTE(mondochiba @ 09/05/07, 03:40 PM) [snapback]272844[/snapback]
with threads set to 1 i sometimes get overall cpu usage of 25% and it renders faster than with 2 threads and cpu usage of 40%

Theres several, and i must say misunderstood facts of multiple cpu's. (this is not just for NUKE it applies for ALL computing)

Hyperthreading: Its not entirely a TRUE processor so if you can fill in the registers of the proc perfectly or near perfectly then the hyperthread s obsolete. Offcourse since the cpu meter THINKS its a real processor it repros it as such in no use while it is 100% in use by the second thread.

Bandwidth: 2 processors isn't better than 1, it is a compormize of what we can do versus what we cant. The problem is that 2 processors gain more overhead because they need to get served data sustained and unfortunately you now have half of what the 1 proc had in bandwidth. But also on top of that theese procs need to spend some of the bandwidth to inter communicate.

Now certain tasks stress different things in the processing so sometimes 1 is actually going to be better than 2 because theres just too much wait / overhead between the 2 procs. This si the reason why therr are clusters and supercomputers side by side.

[flaviomeibach]

there is no hyperthreading on my system so this can be ruled out as a reason.
i am aware of the overhead issue when delegation work to severeal cpus/cores instead of one and that it depends on the task, i still don't think, this is the only problem

since i work alot with different cpu-heavy applications, mostly 3d, i am used to monitoring cpu and memory usage while rendering, just to be aware of any bottlenecks. memory is the problem most of the time, paging even small files to the disc forces the cpu into unneccessary idle-times.
also, not all apps use multithreading equally well (if they do at all)

that's why i would like to know more about the threading in nuke. i think that for nuke to work entirely in float and still be that fast, there must be some clever coding behind it and maybe, some of it just does not go along too well with some of the more rusty parts of xp.

but i might be totally wrong, of course. i am not a td, so this isn't exactly my field of expertise.