Faster Processing For Backtesting?

Sparohok · Dec 20, 2005

A RAM drive is not a good idea. The OS does filesystem buffer caching in RAM, that's more or less a RAM disk for the data you use frequently.

You're on the right track though. With the CPU usage you posted you are clearly IO bound.

What will help? Assuming you can't do anything to optimize the code or simplify the dataset:

1) More system RAM. With enough buffer cache you will no longer be IO bound. Also helps to shut down programs that use a lot of RAM.

2) If that doesn't help enough, faster IO. A RAID0 with a few 10k drives would probably help.

Martin

SideShowBob · Dec 21, 2005

I still doubt a ramdrive will help....but it's worth a shot. I used to use one from here:

http://www.ramdisk.tk/

SSB

fletch2 · Dec 21, 2005

Quote from Norm:

Hello,

I am new to using TradeStation for backtesting. Yesterday I ran an optimization that took about 20 hours to complete. I can foresee much longer optimization later.

I have a dual core Pentium D machine with 2 gigs of RAM. While running, I noticed that one process was 30% busy and the other was 60%.

Can I make the computer work harder, i.e., 100% on both processors, to speed up the test?

Does anyone make a RAM drive, so that I can eliminate disk accesses to gain more speed?

Thanks,

Norm

You will get far more dramatic speedup by rewriting your strategy in a compiled language like C++, than by trying to wring more performance out of your hardware.

It is a lot of work, but if you are serious about accelerating your calculations, it's the only way to go. By tweaking hardware, you might, best case, get 50% speedup. By re-writing, you could probably get a factor of 5 immediately and 10 or 20 with some work.

Cheers,
Fletch

truncheon · Dec 21, 2005

Optimax has solved this problem. Why spend hundreds on hardware or weeks on code to get something that will never even approach what this <$100 service does?

PitchBlack · Dec 21, 2005

Quote from Norm:

Hello,

I am new to using TradeStation for backtesting. Yesterday I ran an optimization that took about 20 hours to complete. I can foresee much longer optimization later.

I have a dual core Pentium D machine with 2 gigs of RAM. While running, I noticed that one process was 30% busy and the other was 60%.

Can I make the computer work harder, i.e., 100% on both processors, to speed up the test?

Does anyone make a RAM drive, so that I can eliminate disk accesses to gain more speed?

Thanks,

Norm

http://www.tomshardware.com/2005/12/05/hyperos_dram_hard_drive_on_the_block/

fletch2 · Dec 21, 2005

Quote from truncheon:

Optimax has solved this problem. Why spend hundreds on hardware or weeks on code to get something that will never even approach what this <$100 service does?

Optimax appears to be just a genetic alg optimizer built on top of TS.

(BTW, their advertising is disingenuous, comparing exhaustive search to the genetic alg. That's apples to oranges.)

It might be a reasonable solution for the OP if his complexity is in the expanse of the parameter set, not in the strategy itself. If the complexity is in the strategy, then a genetic alg won't help him. If it's in the parameter space, it probably will.

I bet you could write a swarm optimizer in less than day that would compare favorably to the genetic alg, and have the combo still test out factors of 10-20x the TS based solution + Optimax, due to the speedup in the strategy eval phase. But, all of that requires skills that may not be worth acquiring for OP.

Cheers,
Fletch

newbunch · Dec 21, 2005

I tried some genetic algorithm optimizer for Excel (don't remember which ones) and found them to be much slower than Excel's solver. But it depends on the type of problem you are trying to solve.

psytrade · Dec 23, 2005

try amibroker and use io.zip to optimize multiple parameters. shouldnt take long.

nononsense · Dec 25, 2005

Quote from fletch2:

You will get far more dramatic speedup by rewriting your strategy in a compiled language like C++, than by trying to wring more performance out of your hardware.

It is a lot of work, but if you are serious about accelerating your calculations, it's the only way to go. By tweaking hardware, you might, best case, get 50% speedup. By re-writing, you could probably get a factor of 5 immediately and 10 or 20 with some work.

Cheers,
Fletch

Basically you are right.
On rewriting: it is often not that much work if you know what you are doing. Mostly, speed has to do with only a few segments of your code. Rewrite these, don't bother with the rest.
It is difficult to give any definite figures, but typically this may concern less than 5-10 percent of your code. Profiling is a good idea.