OpenCl and the tools for it. Reviews and impressions. - page 13

New comment
 
Urain:

Opening price testing in MT4 and MT5 are completely different modes. In MT4 (in your example) will be testing at open H1, whereas in MT5 at open M1, feel the difference by a factor of 60.

I feel the difference in that I don't need M1 testing for an EA working on H1 timeframe by open prices. Maybe you need all kinds of crap and crutches that MT5 developers packed into the platform, but for me and others who deal with auto-trading all this crap is completely unnecessary and useless.

Urain:

In general your whole post looks like a phrase: Ferrari is rubbish, it has low landing, so it will not pass on our roads.

At that, the critic does not care that the car was not designed for our roads.

Maybe a Ferrari is not rubbish, but such car is not suitable for our off-road conditions. That's why I'm doing without Ferrari now.

For the same reason most people are in no hurry to switch from MT4 to MT5, because the platform is not developed for autotrading, but only for owners of expensive hardware, necessary in the field of computer games. The developers confused the marketing positioning: gamers and traders - that's the result.

 
Reshetov: The platform was not developed for autotrading, but only for the owners of expensive hardware needed in the field of computer games.

Yura, don't twist things around. Cloud really does increase the speed of optimization, even if the local machine has no discrete graphics card.

Even if it has a single-core Celeron 420 (which not only supports SSE2, but also SSE3). This is definitely a qualitative difference from the quad.

Well, those who need one will buy one, of course. But you can do the same with Quadruple and use OpenCL (but, of course, it's much more painful).

 
Reshetov:

...For the same reason most people are in no hurry to switch from MT4 to MT5, because the platform was not developed for auto-trading, but only for owners of expensive hardware, required in the field of computer games. The developers confused the marketing positioning: gamers and traders - that's the result.

Agreed. In addition, complication of coding approaches will, I think, scare away a large part of traders who are happy to be creative without OOP.

The popularity of MT4 was built largely on the prevalence of solutions and accessibility for learning the platform "from scratch". Five has raised that bar, and as a result it may be out of reach for many.

 
Reshetov:

Tried it. The optimization speed compared to MT4 is only depressing. MT5 won't dream of the optimization speed of current builds of MT4 yet.

Renat, you are blatantly wishful thinking. I do not argue that MT5 supports various bit architectures and uses all local cores. Another question is how does it use it? I.e. it is possible to use all of them, and at the same time it is absolutely useless.

Renat, have you personally tested this distributed computing network? The impression is that you have bought into the advertising your company is spreading.

Unlike you, I personally test the cludes. I have also been developing, testing and planning this project for many years.

So leave the claims of my ignorance in this area alone.


About OpenCL hardware-independent - this already should be placed in the Humor section. Tell that joke to traders who have on-board video cards without Cuda. This card is more than enough for trading, and for watching quality video and pictures. But for computer games this hardware-independence is not enough. Trading shouldn't be confused with computer games.

About the platform dependence of the GPU you have already proved your error. You don't need to drag it out in public time and time again.

Just because you don't have a GPU, doesn't mean the rest of us don't have one. GPUs are already everywhere and there will only be more to come.


My experiments with Clouds Network gave me these results:

1. There are not 2000 agents on Clouds Network. There are 1500 at most.

2. When Clouds Network is connected, at best of all the agents, only 200 are connected, at worst none.

3. Distributed computing works jerkily, i.e. first autorized mode appears, from which many networks never leave. Then one network connects. It receives a bunch of tasks, engages only part of the agents, quickly processes these tasks and shuts down, going to finished mode for a long time.

I don't say "you haven't tested in practice" for nothing. Maybe you pressed the button a couple of times with fright and on those few seconds the tests were over. After all, that's exactly what happened - otherwise you would have clearly seen the mains warm-up process within the first 30 seconds of start-up.

Here are the results I just ran - there are 7 260 agents from Clouds working on the tasks:

Each pass takes about 25 seconds:

2012.02.06 16:27:51     MQL5 Cloud USA  pass 45855 returned result 103167.28 in 26 sec (PR 83)
2012.02.06 16:27:51     MQL5 Cloud USA  pass 46195 returned result 108641.04 in 25 sec (PR 88)
2012.02.06 16:27:51     MQL5 Cloud USA  pass 46201 returned result 111343.52 in 25 sec (PR 101)

You can see that 55,000 passes in the cludes took 9 minutes, the total expected time is 3 hours and 30 minutes for 1,276,290 passes (each with a 25 sec average).

If we count the time of the local agents (8 of them), we get that they will complete 1,276,290 passes in 25 * 1,276,290 / 8 = 3,988,406 sec = 66,473 minutes = 1,107 hours = 46 days

Here's the real acceleration: 1,107 hours / 3.5 hours = 316 times.


You can see the number of agents on the MQL5 Cloud Network official website. This data is not updated in real time and does not show the upper limit of agents, as the network is sleeping most of the time and rises from sleep only when a large number of orders appear.





Compared to those farms that handle 3D graphics, including free ones, Clouds Network is pathetic rubbish. It's not about any kind of revolution. Everything is made extremely ugly.

The prices are really ridiculous and everything else looks very sad and even depressing.

This is your opinion, built solely on attempts to belittle by making blatantly erroneous statements without practical testing.

Also, you are talking about trading platforms which is scary. You're very far from understanding the business of developing trading platforms.

You make yourself look pathetic.

 
MetaDriver:

Yura. You're right. That's the way it is for now.

But you still have to learn mql5... :) Because the situation is about to change. The developers have already got used to it. And if they have, they will do it. They have also developed the MT4 tester, right?

He is fundamentally wrong.

He hides some good aspirations behind such a screen of false claims and nonsense that it leaves no chance of taking him seriously.

 

When someone gets the idea "the cludes are rubbish", think about what it is like to put 10,000 - 100,000 agents to work at the push of a button.

Not to mention the fact that someone planned, calculated, designed, launched, secured the interests of sellers/buyers and enabled any trader to push a button and get results.

The technical implementation of the project, its full integration into the infrastructure, the resolution of financial issues and the release to the mass market is a true revolution in the optimization of trading robots. No one in the world has done this but us.

 

Since the numerical range of both inputs and weights is very narrow in the grid, I decided to try rewriting the test, replacing double with float.

I am quite surprised by the result - it became slower. Why is float 3.9 times slower than double (in my tests)?

Compiled test #7 in the attachment is 3.9 times slower than test #4.

Files:
testud7.zip  11 kb
 
joo:

Since the numerical range of both inputs and weights is very narrow in the grid, I decided to try rewriting the test, replacing double with float.

I am quite surprised by the result - it became slower. Why is float 3.9 times slower than double (in my tests)?

Compiled test #7 in the attachment is 3.9 times slower than test #4.

Since 100% of the code is executed inside the DLL (you can see it in the code), the question should be addressed to the DLL author. The question has nothing to do with MQL5 itself.

  int endTime=0;
  int startTime=(int)GetTickCount();

  for (int k=0;k<Iteration_P;k++)
    {
     MLP(0,InLayer_P,Weight_P,OutLayer_P);    // это вызов DLL
    }

  endTime=(int)GetTickCount()-startTime;
 
Renat:

Since 100% of the code is executed inside the DLL (it can be seen in the code), you should address your question to the DLL's author. It has nothing to do with MQL5 itself.

Of course, the question has nothing to do with MQL5. The question was of general nature, so to speak.

The author of the dll is me. The source code of the dll is here.

 
OnGoing: In addition, the increasing complexity of coding approaches will, I think, discourage a large part of traders who are happy to be creative and do fine without OOP.

Who forbids you to code in 5 without OOP?

OOP is just a feature of MQL5, not a mandatory requirement. If you have a really heavy project, it will be difficult without OOP. But for light projects the usual, procedural style of coding is enough. The differences from quadruplets are in fact small.

1...67891011121314151617181920...29
New comment