Selecting a computer configuration (expert optimisation a priority) - page 4

[Georgiy Merts]

Nikolay Demko:

Once I was running 6 terminals for a valk-forward, each terminal was counting its period, and then I used excel to put everything together in a script.

So the annual history of a very heavy EA, with all real ticks, has been calculated by cloud in two runs of 15 minutes (more precisely, if I may say so with 6 clouds, I have run 6 terminals).

And all this joy cost 10$.

My machine with 8 threads and 16GB RAM, would have worked this field 2-3 months (if not smog).

And the point in real ticks? There won't be more of them anyway... In my opinion, OHLC on 1M is a great mode.

Or was it a scalper ? Then you can still say something...

I don't see the point of going below M5. Therefore, I try four variants - M5, M15, H1, H4, plus time restrictions, and most experts in testing usually choose H1, a bit seldom M15, quite seldom H4, and very rarely M5 - only few Expert Advisors out of more than 300 selected, and those Expert Advisors are not very successful.

Of course, if there is $10 daily per cloud - then it must be used. But, in my situation, it is unrealistic.

[Mykola Demko]

Georgiy Merts:

What's the point of real ticks anyway? There won't be any more anyway...

Or was it a scalper ? Then you can still say something...

I do not see the point of going below M5. Therefore, I have tried four variants - М5, М15, Н1, Н4, plus time constraints, and most Expert Advisors have chosen Н1 more often, М15 less often, Н4 much less often, and М5 - only few Expert Advisors out of more than 300 chosen, and those Expert Advisors have not so much success.

It is understandable, if there is a daily $10 per cloud - then it should be used. But, in my situation, it is unrealistic.

Scalper.

I do not spend $10 every day. I can test something in course of time on local cores and if I see that it will take me more than 2 hours on local core, then I spend a dollar and spend 10 minutes to develop it in a cloud.

10$ is the final optimization, when all the codes are debugged, all pre-checked, no bugs, and you need a confirmation on a large sample of data. Well, this is one or two times a month at most (faster writing unfortunately does not work).

[Georgiy Merts]

Nikolay Demko:

Scalper.

Then yes.

Alas, with scalpers, it just doesn't work out. I cannot even make a "tester grail".

So, my timeframe is H1.

[Vladislav Andruschenko]

Nikolay Demko:

Scalper.

I do not spend $10 a day. Something in the course of testing I can check on local cores, if I see that the local will be more than 2 hours, then I spend a dollar and make 10 minutes in the cloud.

10$ is the final optimization, when all the codes are debugged, all pre-checked, no bugs, and you need a confirmation on a large sample of data. Well, this is one or two times a month at most (faster writing unfortunately does not work).

Does the cloud already support real ticks? Sorry for asking.

[Mykola Demko]

Vladislav Andruschenko:

Does the cloud already support real ticks? Sorry for asking.

Yes, it does, I checked it on my locals, the same passages match with the cloud.

I also unloaded ticks from the tester and compared them to real ticks accumulated.

So everything works. Although, on the annual history there are a couple of three days with broken ticks.

It's better to download ticks of needed period, and run through the check by script.

Or at least unload it to Excel and look at the chart.

[Vladimir M.]

Georgiy Merts:.............. the question of the purchase of a new system unit arises

Just a reminder, there's useful information and discussion here, on topic.

What video card to choose for computing? ( not mining ) https://www.mql5.com/ru/forum/231043

Since I'm looking to buy a new computer myself, I'm impressed by the following:

Renat Fatkhullin 2018.03.07 12:15 #1 RU

"...Since in finance you have to do the maths, only AMD is the choice.

Nvidia has deliberately slashed the speed of working with double numbers in mass.... cards"

[Sergey Chalyshev]

Georgiy Merts:

Greetings, everyone.

I'm getting more and more stressed by my old system unit. After all, for almost 10 years it is not only obsolete, but it is already limiting me considerably.

So the question arose about getting a new system unit (I have a monitor and it suits me fine).

So.

Aim - maximum performance while optimizing the TC.

Budget - $500.

I'm partial to Intel's CPUs (I'm looking forward to i5 six-core units) and Asus boards (I really like their ability to control fan speeds).

Gaming capabilities are completely unimportant. The last thing I played this year was Age Of Empires (the very first one, from 1997). What are the newfangled toys out there now - not even aware of it.

Lower power is desirable - I like the quietness, especially when I don't need the performance. That's why it's important that the motherboard knows how to control all the fans in the system.

What do you suggest, colleagues?

MQL is not friendly with AMD. Intel i7 and 16 gigs, the drive is not important but ssd is better.

[Fast235]

Sergey Chalyshev:
MQL is not friendly with AMD. Intel i7 and 16 gigs, the drive is not important but ssd is better.

What technologies does Intel use? Before Ryzen, AMD should not be considered.

[Maxim Dmitrievsky]

Build it yourself, it will be 20-30% cheaper

AMD Ryzen 3 or 5 with integrated graphics (around 10kr), 8g ddr4, cheap motherboard, cheap ssd, block and case (for the rest)

budget varic not a gamer, all the work will fly

Optimization is an evil thing, buying a computer for it is... well... yeah.

[fxsaber]

1. Lots of videos detailing how to build a really budget super-computer - dozens of cores.
2. Again, there's plenty of information on how to make a RAM drive at such a level that even the operating system will load on it. This solution is in some cases much stronger than an SSD.
3. Significant speedup can sometimes be achieved by algorithmic optimization. Use TC-version for the Tester, where OOP-overheads, error handlers, preparation of strings to unworkable in the Optimization mode, etc. are minimized. Combat mode of EA (with all the frills) in the Tester is not for Optimization, but for self-testing on a single run.
4. Adjacent runs are very slightly different, so caching is able to still speed up performance.
5. Topics 65 and 66.
6. Custom symbols can speed up Optimisation sometimes by several orders of magnitude. But this mainly applies to real ticks mode. But the crosses will also work.
7. If you are lucky and the weak point is not a lack of a working TC, but slow Optimization only. Then you can write a primitive tester on the pluses with all sorts of caching and integer arithmetic. I used to do it, achieving hundredfold acceleration compared to MT4 in bar test mode.