Thanks for the update. Are you also building Julia with MKL also then?

On Thu Jan 08 2015 at 4:19:48 PM Joshua Adelman <[email protected]>
wrote:

> Hi Jiahao,
>
> Just a small note - based on your comments in the thread you mentioned, I
> ended up changing my test to just multiply ones to avoid over/underflow.
> Those are the results that are now in that notebook, so that shouldn't be
> an issue in the plotted timings. On the python side, I'm using Numpy 1.9
> via the Anaconda distribution built against MKL.
>
> Josh
>
>
> On Jan 8, 2015, at 4:15 PM, Jiahao Chen wrote:
>
> > As Stefan wrote, all you are really doing with larger matrix tests is
> testing the speed of the different BLAS implementations being used by your
> distributions of Julia and NumPy.
> >
> > As I wrote in the other thread
> >
> > https://groups.google.com/d/msg/julia-users/Q96aPufg4S8/IBU9hW0xvWYJ
> >
> > the Vandermonde matrix generation is significantly faster for me in
> Julia than in Python (numpy using reference BLAS).
> >
> > Furthermore Vandermonde is not a good test with larger matrix sizes
> since you are basically testing the speed of multiplying things by
> infinity, which may not be representative of typical computations as it may
> incur overhead from handling overflows.
> >
> > For n=10000 I get the following results:
> >
> > Macbook Julia (Core(TM) i5-4258U)
> > +/-1 matrix: 1.22s
> > [1:n] matrix: 1.21s #mostly overflow
> > rand matrix: 2.95s #mostly underflow
> >
> > Macbook NumPy
> > +/-1 matrix: 3.96s
> > [1:n] matrix: 5.04s #mostly overflow
> > rand matrix: 5.46s #mostly underflow
> >
> > Linux Julia (Xeon(R) E7- 8850)
> > +/-1 matrix: 2.18s
> > [1:n] matrix: 1.89s #mostly overflow
> > rand matrix: 4.36s #mostly underflow
> >
> > Linux NumPy
> > +/-1 matrix: 9.38s
> > [1:n] matrix: 10.64s #mostly overflow *
> > rand matrix: 32.30s #mostly underflow
> >
> > * emits warnings:
> > /usr/lib/python2.7/dist-packages/numpy/lib/twodim_base.py:515:
> RuntimeWarning: overflow encountered in power
> >   X[:, i] = x**(N - i - 1)
> > /usr/lib/python2.7/dist-packages/numpy/lib/twodim_base.py:515:
> RuntimeWarning: invalid value encountered in power
> >   X[:, i] = x**(N - i - 1)
>
>

Reply via email to