Optimization on a dual-processor computer - page 4

New comment
 
I have 24 threads 52 RAM on my server mother. MT5 loads both CPU and RAM 100%. I understand that it is possible to take more, but my friend on Chinese motherboards heats up and resets the frequency. And it is dangerous to overheat the computer.
 
Vyacheslav Vorobev:
I have 24 threads and 52 RAM on my server motherboard. MT5 maxes both CPU and RAM at 100%. As I understand it is possible to take more, but my friend has it on Chinese motherboards and resets clock speed. And it is dangerous to overheat the computer.

Well, I'm leaning more and more towards the idea that there's no point in taking any kind of hat from Alick, it's better to get it from DHS or somewhere else, and assemble it. At least I have a place to complain if something goes wrong.

AMD Ryzen 9 3900X

 
Sergei Makarevich:

This2673 v3 CPUis better in terms of performance and tests

Exactly, but the money was already invested in the v3
 
do agents on OPENCL from video cards work in MT5?
 
Andrey Khatimlianskii:

You can run the optimisation yourself in several steps -https://www.mql5.com/ru/code/26132

But there's really no point in going through all that many combinations.

This is the first thought that comes to mind. But several stages is not the same thing. Because sometimes, if you change the parameters slightly, the results change significantly, and important parameters will simply be missed. There is no dependency so that you can find limits in large steps and then give smaller steps in that limit. For this reason, in order to test some hypotheses, the limit of the tester passes should be removed. Whether there is no point or not, only experience can judge. Research sometimes requires an unconventional approach.

 
Aliaksandr Hryshyn:
And who will control what happens on the wind? Cool hardware won't help? Too bad, then what is the point of it? Security will get worse, there are a lot of things the average user doesn't know about. Security is first of all a matter of user classification.
The user's actions are up to the user, he can correct them if he wants. But presence of a hardware backdoor does not depend on the user because it cannot be deleted even if desired. Re-flashing the bios in this case won't help. The baddor is located in the South Bridge. Moreover, nothing appears where a normal user cannot figure it out. A very serious level of security can be achieved by following a simple set of some rules that require a minimum level of hardware and operating system knowledge to understand. I am sure anyone can do this.
 
Sergei Makarevich:

Well, I'm leaning more and more towards the idea that there's no point in taking any kind of hat from Alick, it's better to get it from DHS or somewhere else, and assemble it. At least I have a place to complain if something goes wrong.

AMD Ryzen 9 3900X

by the way, while researching ryzen I found a pretty good deal on the ryzen 7 2700x... For the price of the 9 you can build a complete PC (without the external graphics card), yes there will be less cores/threads, but the cost per core/thread is lower.

Of course, if you need a number crusher for MT, then 2x processors on xeon will probably still be above competition on price/performance, but their future liquidity and using for other tasks, e.g. video processing, is under great question in view of limited set of processor commands and low (relatively modern CPUs) frequency...

PS also was getting closer to this topic from an EA optimization side, but while working on an old office computer somehow the topic has disappeared and the need for a lot of threads/core is not necessary anymore (or is it so far?).

 
pribludilsa:

That's the first thought that comes to mind. But a few steps are not the same thing. Because sometimes, if you change the parameters slightly, the results change significantly, and important parameters will simply be missed. There is no dependency so that you can find limits in large steps and then give smaller steps in that limit. For this reason, in order to test some hypotheses, the limit of the tester passes should be removed. Whether there is no point or not, only experience can judge. Research sometimes requires an out-of-the-box approach.

Divide your full table of over 100,000,000 passes by 100 passes of 1,000,000. You will end up with the same complete table of results (can be glued together programmatically).

 
Andrey Khatimlianskii:

Divide your complete overlap of over 100,000,000 passes by 100 passes of 1,000,000. You will end up with the same complete table of results (you can programmatically glue it together).

How to divide? Find first one parameter, then another, will not work, because there is a correlation of parameters with each other, and the correlation may be in the form of a complex function, the very search for which is rather a false approach. So, this kind of separation of optimisation is a mistake. A more correct way is described in my post, where you can try to find limits of good parameters in huge steps, but since results of insignificant parameter changes can change in leaps and bounds, this approach is not suitable either. One could of course try to work with deriving parameter correlation functions, but why, since such an approach is complicated, when one can do it in a more direct way, simply by doing a slow optimization with a large number of passes. This would seem to require writing one's own tester.

 
And we should also add, if we take up the derivation of the parameter correlation functions, we again encounter the same problem, the possibility of function graph jumps in a very narrow limit. Since there is a correlation of each with each, we cannot first find one dependence, then another and so on, again we will get a huge task that needs to be broken down in huge steps. In this way there will be an accumulation of error, which can be avoided by simply using optimization with a large number of passes. Not to mention that this approach in general is extremely complicated, it seems to me. I am probably not using the word correlation correctly, I mean that all the parameters act together on the result.
12345678
New comment