houdini 11 vs. houdini 9.5

[petz]

i´m currently working on an old project where i had to do quite a lot of ray intersections in sops and vops. houdini 11 is loading the old 9.5 - files without problem but it takes around 10 times longer to calculate the intersections than it did in 9.5. after doing some tests it seems that houdini 11 and 10 are much slower than 9.5 by a factor of 5 - 15.

so my question is - can this be confirmed on other platforms ( other than win 64) or can it confirmed at all, and if yes, does someone know what´s causing this?

simple testfile is attached!

thank´s

petz

perf_test.hipnc

[crunch]

i´m currently working on an old project where i had to do quite a lot of ray intersections in sops and vops. houdini 11 is loading the old 9.5 - files without problem but it takes around 10 times longer to calculate the intersections than it did in 9.5. after doing some tests it seems that houdini 11 and 10 are much slower than 9.5 by a factor of 5 - 15.

Hi,

I timed cooking 10 frames in H11 and in H9.5. I got 5.05 seconds in H11 and 4.9 seconds in H9.5, which is a little slower, but nowhere near 10-15 times slower.

What OS/hardware?

[petz]

hi crunch,

thanks for taking time to do the test. the result is interesting!

when i did the timing questerday i had times as follow for 100 frames.

ray-sop | vops-singlethreaded | vops-multithreaded

houdini 9.5 16.78 sec | 13.78 sec | 13.44 sec

houdini 11 1 min 1.02sec | 58.91 sec | 34.32 sec

houdini 11 is around 5 times slower on the test-scene. if you increase the number of points the performance loss is even more existent and steady rises up to a factor around 10. first i thought this could be somehow related to viewport performance but using hbatch i got more or less the same result. so i repeated the procedure on different machines but houdini 11 was slower there as well.

i´m not sure whats the reason for this, if it has something to do with combination of hardware and os but its a bit annoying to wait 10 hours in H11 instead of 1 hour with H9.5.

so any tip will be welcome ( switching to linux is not an option )

btw.

all test have been done on vista 64, dual xeons, quadro fx 3400 with latest driver installed.

petz

Edited September 9, 2010 by petz

[Macha]

Well, I get this for a typical frame in H11.

Windows xp64, i7.

 Start Frame     2:
   105.98 ms   468 cook    /obj/geo1/point1 (Sop/point)
/obj/geo1/point1: Memory Usage Delta: 0 (287,150,080)
     0.08 ms   468 cook   /obj/geo1/cache1 (Sop/cache)
/obj/geo1/cache1: Memory Usage Delta: 0 (287,150,080)
     0.72 ms   468 cook   /obj/geo1/mountain1 (Sop/mountain)
/obj/geo1/mountain1: Memory Usage Delta: 0 (287,150,080)
   118.52 ms   454 cook  /obj/geo1/ray1 (Sop/ray)
/obj/geo1/ray1: Memory Usage Delta: 0 (287,150,080)
     0.01 ms   477 port   background (persp1)
     0.00 ms   477 port   pre-render (persp1)
     0.47 ms   draw          /obj/geo1/ray1
     0.05 ms   draw          /obj/geo1/mountain1
     0.21 ms   477 port     inter-object (persp1)
     0.09 ms   477 port    normal-quality (persp1)
     0.20 ms   477 port   beauty pass (persp1)
     0.07 ms   477 port   unlit (persp1)
     0.00 ms   477 port   handles (persp1)
     0.01 ms   477 port   post-render (persp1)
     0.00 ms   477 port   foreground (persp1)
     0.04 ms   477 port   name (persp1)
     0.10 ms   477 port   statename (persp1)
     0.16 ms   477 port   other (persp1)
   End Frame     2:    235.72 ms       4.24 Hz
  Unaccounted Time:      9.02 ms 
   Average of  229:    650.28 ms       1.54 Hz

(how do I get total cooking time over several frames?)

Edited September 9, 2010 by Macha

[hopbin9]

@Macha - How did you generate that report?

I'd really like to time how long it takes to render a frame using the render window. I'm a bit of a noob and can't figure out how to find the render time when doing a single frame.

[malexander]

That's the Performance Monitor (Windows>Performance Monitor). It doesn't monitor renders, just cooking, drawing and some other Houdini-specific tasks.

[daniel.phillis]

Ive still not found an easy way to get render times for a single frame.

ive heard people talk about python pre-frame and post frame scripts - but i dont know python...

though that shouldnt be too hard...

last time i looked at the finished rendered bitmap's creation time and compared it to the next bitmaps creation time

...eech !

Edited September 10, 2010 by daniel.phillis

[petz]

macha, is there a chance that you could run the testfile on houdini 9.5 as well, i would really like to know how houdini 11 performs compared to 9.5 on your system.

this problem makes me quite nervous as long as i have no idea what´s causing the trouble.

thank´s,

petz

[Macha]

On 9.5 I get this:

Start Frame     2:
    66.65 ms   111 cook    /obj/geo1/point1 (Sop/point)
/obj/geo1/point1: Memory Usage Delta: 0 (306,819,072)
     0.62 ms   111 cook   /obj/geo1/cache1 (Sop/cache)
/obj/geo1/cache1: Memory Usage Delta: 0 (306,819,072)
     0.60 ms   111 cook   /obj/geo1/mountain1 (Sop/mountain)
/obj/geo1/mountain1: Memory Usage Delta: 0 (306,819,072)
    43.06 ms   111 cook  /obj/geo1/ray1 (Sop/ray)
/obj/geo1/ray1: Memory Usage Delta: 0 (306,819,072)
     0.05 ms   117 port   background (persp1)
     0.05 ms   draw        /obj/geo1/mountain1
     0.35 ms   draw        /obj/geo1/ray1
     0.11 ms   117 port   transp (persp1)
     0.15 ms   117 port   inter-object (persp1)
     0.00 ms   117 port   handles (persp1)
     0.00 ms   117 port   foreground (persp1)
     0.01 ms   117 port   name (persp1)
     0.05 ms   117 port   statename (persp1)
     0.17 ms   117 port   other (persp1)
   End Frame     2:    141.50 ms       7.07 Hz
  Unaccounted Time:     29.64 ms 
   Average of    5:    153.61 ms       6.51 Hz

It seems mountain and point sop are faster.

Edited September 10, 2010 by Macha

[Pazuzu]

Its a shame that some operators in H11 are more slow than 9.5.

[edward]

first i thought this could be somehow related to viewport performance but using hbatch i got more or less the same result. so i repeated the procedure on different machines but houdini 11 was slower there as well.

I'm not getting your results. I'm on WinXP x64, Intel i7 (920) 2.67 GHz, 12 GB RAM. I also have the .hip file attached that I used. It's pretty much the same as your file except with some display option changes and converted to non-commercial. How many points were you using when seeing the 5 times speed differences?

$ hbatch RayTest.hip
hbatch Version 9.5.457 (Compiled on 05/03/09)
/ -> time opcook -F -f 1 100 /obj/geo1/ray1
17.7143r
/ ->

$ hbatch RayTest.hip
hbatch Version 11.0.499 (Compiled on 09/10/10)
/ -> time opcook -F -f 1 100 /obj/geo1/ray1
16.4781r
/ ->

RayTest.hip

[petz]

@macha:

its not as much difference as i had but the ray-sop is also almost 3 times slower on houdini 11 as seen in your test.

@edward

that`s strange! i´m getting this results with you scene:

hbatch Version 9.5.371 <Compiled on 02/06/09>

/obj/geo1/ray1 - 11.6789r

/obj/geo1/vopsop1 - 10.7932r

hbatch Version 11.0.491 <Compiled on 09/02/10>

/obj/geo1/ray1 - 39.2324r

/obj/geo1/vopsop1 - 26.9872r

houdini 9.5 ui

/obj/geo1/ray1 - 17.2119999886

/obj/geo1/vopsop1 - 13.7600002289

houdini 11 ui

/obj/geo1/ray1 - 59.3140001297

/obj/geo1/vopsop1 - 33.1549999714

using hbatch obviously is faster but the ratio between houdini 9.5 and 11 is much the same. houdini 11 is around 3.4 times slower in both cases. when i increase the number of points beyond 150000 i come close to a factor around 5.

all in all i´m inclined to think that there is something not working well together between the computer (vista 64, dual intel xeon 5430, 12gb) and houdini 11, but i have no idea what exactly that could be.

anyway thanks for doing the benchmarks

petz

[edward]

Those are really odd results. Are you running in a clean environment? Perhaps try moving aside your $HOME/houdiniX.X directories and make sure that you don't have any Houdini-related environment variables set. Also make sure that you don't have any extraneous programs running.