forum.alglib.net

hello Alglib,
Have been trying the free C++ version and it's ssa routines mainly. Works great, thanks. Would be interesting and useful for me to see if the threading options & Intel MKL inclusion in commercial version would speed it further up. So, wonder if you still have trial version available? Requested it there on the page but didn't hear back anything. Unfortunately don't have corporate email... (but could consider purchasing it still)

thanks, mikko

Dear Mikko!

Your trial request was approved shortly after submission. It seems that automated message from alglib.net was blocked by your anti-spam filter. I will forward you a reference shortly from another account.

Thanks. Hope to receive it.

Thanks for the trial. Your commercial version really reaches impressive results. Your test ended up now to 60.4 GFLOPS in my machine. That's up from 9.8 in free version (or actually 1.9 at first). Looks great!

But I didn't get any speedup in my real application. Good is that It works as earlier even though the added defines, library and dll are in place, and that the integration/compilation went smoothly. I'm using alglib in Windows DLL which is called from MT4 (FXCM forex) environment. Ssa calculation takes now half of the backtesting time, and increases quickly if/when I configure more of that in there.

I guess I need to take somehow a closer look which is now possible with the sources. Thanks.

What settings do you use for SSA - window width? Number of components? SSA with small number of components (tens) and large window widths is mostly bandwidth limited.

I've tried and am trying different from several tens to few hundreds mostly. And the same for components from few to few tens. But I don't really have the theory knowledge so I'm just trying to find something which would work (though such may not even exist). I'm feeding the forecast to neural network (not yours but fann lib) and trying to train that as well.

I tried now also the "own installation of Intel MKL" option as well. Almost the same results: standalone app gives high speedup but again no difference in dll. I'd expect at least some change in behaviour so I need to miss something. Traced the execution and seems that it enters mkl lib, will now still see if that happens also with alglib2mkl.c.

Verified further. Alglib works perfectly. Can see the speedup in the dll if I run the sample test code in there with different n's and worker amounts. Seems that mkl gives boost with quite small n's. There's just something else as I don't get any overall speedup in the application itself. Need to give up or ask from somewhere else. It was exactly the same when I was trying TBB etc. Strange. Would probably need to code it all with c etc and leave mql4 and mt4/5 completely.

Also after some my own fixes SSA doesn't take half but around 23% when using on three different timeframes (with different window and component sizes). I had neural stuff included accidentally earlier and calculated unnecessarily the forecast too often. One thing I thought was that maybe it would be nice if one could replace the latest data point in input stream instead of having the possibility to just add data points. Otherwise the appendpointandupdate stuff in SSA was exactly what I was looking for.

Thanks.