Super User is a question and answer site for computer enthusiasts and power users. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

EDIT: For anyone who stumbles upon this in the future: Imagemagick uses a MP library. It's faster to use available cores if they're around, but if you have parallel jobs, it's unhelpful.

Do one of the following:

  • do your jobs serially (with Imagemagick in parallel mode)
  • set MAGICK_THREAD_LIMIT=1 for your invocation of the imagemagick binary in question.

By making Imagemagick use only one thread, it slows down by 20-30% in my test cases, but meant I could run one job per core without issues, for a significant net increase in performance.

Original question:

While converting some images using ImageMagick, I noticed a somewhat strange effect. Using xargs was significantly slower than a standard for loop. Since xargs limited to a single process should act like a for loop, I tested that, and found it to be about the same.

Thus, we have this demonstration.

  • Quad core (AMD Athalon X4, 2.6GHz)
  • Working entirely on a tempfs (16g ram total; no swap)
  • No other major loads


/media/ramdisk/img$ time for f in *.bmp; do echo $f ${f%bmp}png; done | xargs -n 2 -P 1 convert -auto-level

real        0m3.784s
user        0m2.240s
sys         0m0.230s
/media/ramdisk/img$ time for f in *.bmp; do echo $f ${f%bmp}png; done | xargs -n 2 -P 2 convert -auto-level

real        0m9.097s
user        0m28.020s
sys         0m0.910s
/media/ramdisk/img$ time for f in *.bmp; do echo $f ${f%bmp}png; done | xargs -n 2 -P 10 convert -auto-level

real        0m9.844s
user        0m33.200s
sys         0m1.270s

Can anyone think of a reason why running two instances of this program takes more than twice as long in real time, and more than ten times as long in processor time to complete the same task? After that initial hit, more processes do not seem to have as significant of an effect.

I thought it might have to do with disk seeking, so I did that test entirely in ram. Could it have something to do with how Convert works, and having more than one copy at once means it cannot use processor cache as efficiently or something?

EDIT: When done with 1000x 769KB files, performance is as expected. Interesting.

/media/ramdisk/img$ time for f in *.bmp; do echo $f ${f%bmp}png; done | xargs -n 2 -P 1 convert -auto-level

real    3m37.679s
user    5m6.980s
sys 0m6.340s
/media/ramdisk/img$ time for f in *.bmp; do echo $f ${f%bmp}png; done | xargs -n 2 -P 1 convert -auto-level

real    3m37.152s
user    5m6.140s
sys 0m6.530s
/media/ramdisk/img$ time for f in *.bmp; do echo $f ${f%bmp}png; done | xargs -n 2 -P 2 convert -auto-level

real    2m7.578s
user    5m35.410s
sys     0m6.050s
/media/ramdisk/img$ time for f in *.bmp; do echo $f ${f%bmp}png; done | xargs -n 2 -P 4 convert -auto-level

real    1m36.959s
user    5m48.900s
sys     0m6.350s
/media/ramdisk/img$ time for f in *.bmp; do echo $f ${f%bmp}png; done | xargs -n 2 -P 10 convert -auto-level

real    1m36.392s
user    5m54.840s
sys     0m5.650s
share|improve this question
Another thing to remember about parallel processing is that there is an overhead associated with managing the threads and ensuring the program runs without deadlocks etc. – ChrisF Jul 28 '11 at 20:50 seems to indicate that this is a problem with imagemagick's SMP code. – zebediah49 Jul 28 '11 at 21:37
up vote 4 down vote accepted

How big are the files you wish to convert when compared to your L1 cache? Your L2 cache?

Without a better look inside, I would suspect cache contention is causing your CPUs to idle while they wait for data to be re-cached because the other process(es) keep kicking important things out of the fast memory.

See also this answer I gave on Stack Overflow.

share|improve this answer
Athlon II X4 processors have 512K of L2 cache dedicated to each core and no L3, so it's not a simple cache contention problem. I suspect there are some bugs in the code somewhere that cause clashes on the memory bus, inefficient locking or unnecessary cache line invalidation. You might be better off using a multi-threaded version of ImageMagick instead. In fact, the version you have should be multi-threaded already, so the delay could actually be caused by the threads competing with each other. – Jul 28 '11 at 21:17
dmidecode says I have 512K of L3, no L2, and 128K of L1, but close enough (also, demidecode has been known to lie). The images in question are 3.1KB(54+32*32*3 bytes), which may be related. Convert -version reports Features: OpenMP. I will retry with a set of 769K images. – zebediah49 Jul 28 '11 at 21:32
::sigh:: Per-core caches kinda kill this idea. – dmckee Jul 28 '11 at 21:40
Possibly not, if ImageMagick is spawning 4 threads per copy of itself-- that means they can get in each other's way on every core... I added an example with larger images that works the way I would expect. – zebediah49 Jul 28 '11 at 21:50

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .