Any good hardware guys here?

Hey everyone, I'm getting around to replacing one of my older servers. I had a peek around the dell website, and came up with this:

http://www.dell.com/us/en/business/s...10&s=bsd&cs=04

Pick the option on the right, and then change the options to the lowest processor with HT and Turbo, 8GB ram (4 x 2gb) and minimum everything else (HDD, software, etc). Total cost: $575 with coupon

Up until now I've built all of my servers, but after seeing that I could get a computer which is 80% the power of my most powerful computer for half the price. Does anyone with experience with dell/hp/manufactured servers have an opinion on this?

I was thinking of installing VMWare ESXi 4.0 on this, and have 8 VMs each running one instance of MT4.

Dell have only good quality servers, nice cumpute power/energy cost ratio. In my company we have ~20 dell servers - working almost 2 years without hardware fail.

For what purpose you have so many MT4 instances? For fail issues?

Primarily testing/optimization for now. Metatrader seems to take a max 1GB of ram anyway. It'll be a lot easier for me to do some optimization runs. Once I'm comfortable with Dell servers as a stable platform, I'll start buying rackmount versions of this and get them properly colocated.

My current 'IT Policy' is to have several cheap servers, each one is redundant. My home-built servers cost me less than $1,000 each. They're located in the closets/datacenters of family and friends, throughout the US. It's definitely cheaper than colocating.

[edit]

I'm still trying to leverage GPUs using openCL/openGL and move the neural network processing into that highly parallel environment. Once I can get this production-ready, I'll almost definitely have to build my own servers.

For GPU processing consider using AMD/ATI GPU, maybe in 3D Games they are weaker than nVidia but when we talkin' about FLOPS/price they are best. Geforce 480: 2017 GFlops/450$ = 4.4GFLOPS/$ with AMD Radeon 5970 4640GFLOPS/550$ = 8.43GFLOPS/$. Of course when we take power consumption and heat problems of GF480GTX.. AMD/ATI is winner.

Problem is ATI's implementation of GPU-based computing isn't as flexible as NVidia's, also, NVidia made a specific GPU called the Tesla which is made specifically for this.

Tesla right now is not best choice when we have new GF480 and ATI/AMD 5970. Just take price, compute power and compare to new GPU's. They are a lot better. ATI is nice optimized for OpenCL so I think it is quite good interface for programming.

I'm not fan of ATI/AMD but i like to use cheap and good hardware. Right now I don't know better product for GPGPU.

Nvidia has only old versions of tesla in markets (ex. S10xx). In specifications new Tesla M2070 will have 1TFLOP in single precision, normal GPU size and price not less than 1,000 $ so anyway Radeon is still best choice.

I agree speed is important, but the computing power of the newer gfx cards is rendered moot if the data can't fit in the available RAM.

How many cash you want to spend for buyin' GPU? Do you have any experience with programming on so many processors?

Programming GPUs is serious shit. You really need to understand their memory model if you really want to get that TFLOPS performance. Use them in a naive way and it will be slower than a CPU.

I'm sure you know they can't be used directly from MQL. You could use DLL dispatch, but it would be harder in my opinion.

Me, I've decided that my performance requirements are not yet that high to invest in GPU computing now.

You can watch this series if you want to know what's really in front of you if you go this way:

Yeah you are right, cores synchro is one of hardest problem, but there is a lot more. In example you need to lock access to part of memory for other cores. There is few types of memory per Core (core mem, group mem, shared mem, computer mem). But with time, and practice programming nice optimized apps is not so terrible.

Probably thread author writing his ANN in C/C++ (there is no way that ann generator can work good in forex). I think when we want to use GPU for processing our computer station need to be in our home

Btw. Nice video

Cost is almost moot for me. I have no qualms dropping 50k to build a bleeding edge computer for my trading--I can write it off as a business expense--the only requirement I have is that it works reliably and pays for itself within a year.

Adal, I have a program which writes NNs for me. I feed it the input layer and output layer, the program then determines the connections, weights, and hidden layer. The problem is it takes my most powerful computer (dual 6 core opteron) about 5-6 seconds per tick. I can't have that. I'm undecided between buying a set of blade servers and running a cluster out of my basement/datacenter, or diving into the world of GPU-based computing.

Well, there is no need to pay the "enterprise tax" and buy a Tesla yet.

You can use you own desktop video card for prototyping a GPU-based NN. I did try using OpenCL for tick filtering and converting ticks to candles, and while it worked and it was 10 times faster than the CPU, which after more optimization could go as high as 100 times faster, it simply does not justify the time I spent on implementing this. I do not need this kind of speed at the moment.

5-6 sec/tick is quite a lot. Not even a GPU setup will save you here You need to eliminate your bottlenecks, whatever they are.

Ronald I think that ANN generators are bottleneck itself. Building ANN is for me very complex, I try to use probably every possible configuration with few types of inputs, few architectures, activation functions etc. to achieve best results. For what purpose you using ANN? Price prediction?

Mode selection.

Input layer is anything I could possibly throw at it. I could put any data as long as I can get it timestamped and match it up with FX data. One particular set of EAs uses 14,000 inputs. Output layer is three things:

 

1. Buy
2. Sell
3. Do Nothing/Unsure

My feedback loop checks which of the two options (buy/sell) would have been the best, and adjusts the weights accordingly based on what type of performance I want the EA to achieve (scalping, trending, breakout, etc).

The connections and hidden layer are determined by the computer, not by me. It's this step which I think is acting as your bottleneck.

Since the EA is basically testing a population of a few hundred NNs simultaneously, I would imagine this qualifies as a highly parallel process.

https://www.forexfactory.com/thread/...2#post12915642