Evaluating CPU cores for optimisation - page 12

 
Aleksey Vyazmikin:

Unfortunately, I am dumb, but there X>=0 may be greater than any other number - there are many combinations - it is impossible to provide everything in the code, and the code will grow by many orders of magnitude due to different combinations.

There are rather many repetitions with one value too. But it will hardly help.

 
I can test a Ryzen 3700x. Am I correct in assuming that with each of the three EAs you need to do a) 16 passes with 8 agents turned on b) 16 passes with 16 agents turned on ?
 
Quantum_Logic:
I can test a Ryzen 3700x. I understand correctly that with each of the three EAs you need to do a) 16 passes with 8 agents turned on b) 16 passes with 16 agents turned on ?

Try it.

 
Aleksey Vyazmikin:

I suggest collecting statistics on the performance of various processors to estimate their effectiveness for the Strategy Tester in optimization mode.

In order to be more objective, I suggest using the Expert Advisor in the"Mathematical Computing" mode which will minimum utilize hard disk and computer memory. If possible, specify the name of the processor, motherboard and frequency of RAM.

This will be a test of the CPU performance on the given algorithms, nothing more. You can "roughly" predict them in advance based on benchmarks: https://www.cpubenchmark.net/high_end_cpus.html

I, for example, have everything in memory and disk mainly (memory just isn't enough and everything gets swapped). And here the same xeons can get ahead of modern i7-8700k for the same price, working in 8 channel mode, for example as described here: https://pikabu.ru/story/domashniy_komp_s_servernoy_nachinkoy_6442138

But, that's when we're talking about overkill. I have to both overshoot and pass singles. That's why I would pay my attention to this table:https://www.cpubenchmark.net/singleThread.html and more specifically to AMD 3600X - 3900X (it depends on your wallet). True, almost all desktop comes with 2 memory channels. The 4 channel server solutions, core performance will be expensive.

Also, I think MQ will make common history for N threads and there will be no memory problem.

And I would also like to give an opportunity to write part of the EA in CUDA with subsequent parallelization on video cards, because there is a much more interesting perspective.

 
Andrey Pogoreltsev:

For me, for example, it's mainly the memory and disk that's holding everything back (there's just not enough memory and everything gets swapped).

What kind of CPU and memory do you have? SSD/HDD?

 
Andrey Pogoreltsev:

This will be a test of the CPU performance on the given algorithms, nothing more. You can "roughly" predict them in advance based on benchmarks: https://www.cpubenchmark.net/high_end_cpus.html

For me, for example, it's all about memory and disk, for the most part (I just don't have enough memory and everything gets swapped). And here the same xeons can get ahead of modern i7-8700k for the same price, working in 8 channel mode, for example as described here: https://pikabu.ru/story/domashniy_komp_s_servernoy_nachinkoy_6442138

But, that's when we're talking about overkill. I have to both overshoot and pass singles. That's why I would pay my attention to this table:https://www.cpubenchmark.net/singleThread.html and more specifically to AMD 3600X - 3900X (it depends on your wallet). True, almost all desktop comes with 2 memory channels. The 4 channel server solutions, core performance will be expensive.

Also, I think MQ will make common history for N threads and there will be no memory problem.

And I would also like to give an opportunity to write part of the EA in CUDA and then parallelize it on video cards, because there is a much more interesting perspective.

So openCL is there, paralleling on video cards is already possible.
 
Pavel Verveyko:

What kind of CPU and memory do you have? SSD/HDD?

I have a Ryzen 5 2600 now, I'm thinking of getting something more powerful.

Memory is 4 x 8GB CORSAIR Vengeance 3000

My main drive is a Samsung 970 500Gb NVMe disk, but I bought a 1Gb SSD with SATA-3 for multi-threaded tests since my tester has been raging the hell out of it every time.

 
Maxim Romanov:
That's how openCL is, you can already parallelize it on video cards.

Will it parallelize in the strategy tester? I.e. can I run N optimization threads on OpenCL on one CPU core?

 
Andrey Pogoreltsev:

Will it parallelize in the strategy tester? I.e. can I run N optimization threads on OpenCL on one processor core?

In the tester, the Expert Advisor's internal logic can parallelize on OpenCL, but not the history passes.

 

OpenCL is certainly nice, but only if we are talking about specific tasks to be solved on the developer's hardware, otherwise the requirements for the product increase dramatically. Such tasks could be:

  1. Drawing 3D graphs/models (questionable task)
  2. Neural network calculation
  3. Parallelization of optimization of the Expert Advisor - that would be very cool. If the Expert Advisor could contain OpenCL code that could be paralleled. But for that the strategy tester should be able to run N instances of the EA regardless of the number of processor cores. And of course first you need to create a shared memory with ticks for the instances.
  4. Mining))))
PS. By the way, it is interesting to run OpenCL mining on Cloud machines and see how high computing power can be obtained)