GPU Backtesting & Research Compute in Beta

Hi Jared

Thank you for the feature! Think that a lot of people will benefit from his new feature. I have 2 question though. What happens if all 3 people use the GPU at the same time? Will each user have access to ⅓ of the ram and process power? Furthermore, can the GPU also be used for live trading? 

Have a good day

Lucas

Hi Luca! The RAM and CPU are reserved(dedicated) and there would be ⅓ of a GPU minimum. Live trading isn't done yet but it's on the roadmap. For now, models can be trained in research/backtesting and passed to live via the Object Store.

Best,
Jared

This is awesome. Great work team.

Hi Jared,

Great feature, since I am trading alone the cost is a bit prohibitive. Would it be possible to have a plan where we pay based on the amount of time used in QCC, so that I only pay when I train it. 

Thanks and understood Anup! We're hoping to roll out something like that soon. The full nodes provide a more predictable workload - and provisioning/sourcing them without predictable demand is difficult at the moment. As the supply of GPUs increases, we'll be able to install more for community QCC use. 

Thanks Jared Broad it looks like a great feature,  Are there any guides to writing a GPU back test as opposed to a traditional back test?  

Also how can you select this GPU back testing?  Upgrade back testing node to GPU and then it will be triggered automatically on each back test?  

Once added to your account GPU node shows up as a node in your resource list - you can force the backtest to use it by clicking the node in the list. 

Re: Writing a GPU backtest - many libraries support GPU with no API changes. However, there are things to consider when writing a GPU backtest. The loading time of the model can be several milliseconds, so you want to do this batches. If misused, it can appear slower than a traditional CPU backtest.

I think this justifies a section of docs so we'll draft something up and post it here shortly.

Hi Jared Broad thank you very much is a very exciting development.  

Hi Semurg Enterprises,

We updated the documentation here.

Best,
Derek Melchin

Hi Derek Melchin and Jared Broad thank you very much.  

If running a backtest / training period can the period be split into weekly units and each weekly unit sent sent to a GPU core with results gathered or saved (e.g. to object store) at the end.  

Something similar to data parallelism as described here:

With structure of Quant Connect script as I understand the def Initialize(self) method gets called first and once which sets the backtest start and end data and then the def OnData(self, data) method is called until the backtest period completes; I am not sure how you could go about splitting up the backtest into weekly units and then running in parallel?  

Is there a way via for example the CLI?  Or another method you could recommed?

Thank You 

Can parameters be passes on command line like this → 

lean cloud backtest MyProject –start 2022-01-01 --end 2022-01-31

And if so how can they be pulled into the Quant Connect script?  

Or is it better to adjust the config file in Lean with start and end dates before each run?

And iterate with a Python / CLI script something like this:

import subprocess
from dateutil.relativedelta import relativedelta
import datetime

# Starting and ending months for backtests
start_month = datetime.date(2022, 1, 1)
end_month = datetime.date(2022, 12, 31)

current_month = start_month

while current_month <= end_month:
   start_date = current_month
   end_date = current_month + relativedelta(months=1) - relativedelta(days=1) # to get the last day of current_month

   # Construct the lean cloud backtest command
   cmd = f"lean cloud backtest myProject --start {start_date} --end {end_date}"

   # Execute the command
   subprocess.run(cmd, shell=True)

   # Move to the next month
   current_month = current_month + relativedelta(months=1)
 

With a GPU backtesting node will each call be pushed to a GPU core?  

I think you're mixing concepts of optimization and GPU. I'll see if we can post an example that harnesses the GPU fully. 

This is great! Research and training my model was super quick, but now I'm incentivized to wait/do more research for the rest of the pay period before changing my subscription. Is there any chance of being able to switch from a Research to a Backtesting GPU without being charged again?

I realized that's effectively what happens, as the billing date just changes and there's a pro-rated adjustment 😅