Fascinating stuff. It's quite beyond my capabilities to hand code all my trading systems in C (had a hard enough time learning AFL). I can't even imagine finding the time to be proficient enough to code my own backtester!
But it's great to know that it's possible. Maybe a workable solution will one day trickle down to the rest of us. I am certain I can make a major improvement in one of my trading systems if I can run this massive optimization I have been drooling over. But the estimated time to completion of the AA is a little over four years! --- In [email protected], "dloyer123" <[EMAIL PROTECTED]> wrote: > > --- In [email protected], "ozzyapeman" <zoopfree@> wrote: > > > > Would AB really run any faster on this PC, or would one have to > design > > some kind of hack that allowed AB to take advantage of the extra > > processors? > > > > My limited understanding was that to get AB's backtesting to run on > > NVIDIA graphics cards, some kind of CUDA kernal had to be designed - > > and that was a huge job that nobody took on. There was a thread > about > > that a few months back. > > > > > Yep, it can be done, it takes more than a little hacking, but yes it > is fast. The bandwidth and computing power available is massive. > > It also takes a lot of low level coding to make it work. I had to > write my own backtester and hand coded my system in C. > > In my application, a portfolio of 1,000 or so symbols with a year or > so of 5 minute bars takes about 25 seconds in AMI and only 5-6ms on > the GPU. Brute force searches work just fine. > > But much of that speed up is due to the use of hand optimized code. > Specifically, it avoids the very large overhead AMI has in each > optimization step, which amounts to most of the execution time. > > Instead, I was able to download the price bars into the GPU once, > then run multiple passes over the data with different parameter > values. Very fast. > > TJ will brag about how fast AFL execution is, and it is faster than > other general tools, but I dont think that he has sat through many > multi day walk forward runs. > > If he did, he would not be so wedded to rebuilding price bars on each > optimization pass and find a way to do it just once. Also, there are > faster ways to handle the signal list than to materialize the full > set of buy/sell/short/cover/price arrays. > > For example, running the code I wrote for the GPU on a laptop without > a GPU, using the GPU emulator, running on a single CPU core, was > still about x10 faster than in AMI alone. > > But, dont underestimate the utility of a high level tool like ami. > Sure, it is slower than hand coded C, even on just one cpu, but a > heck of a lot easier to work with. There is always a performance hit > for a high level tool. > > I dont bother writing C code unless I need to optimize 3 or more > parameters and then, I try real hard to knock it down to improve the > system walk forward efficiency. >
