Skip to content

🐚 Starting Launchers and Running Sessions

Selecting and Starting Launchers

Click on your project and select one of the available launchers to create a new session. Note that each launcher has multiple configurations of CPUs/GPUs and memory, and that these vary by software. image

Select your software, name the session, and choose the CPU/GPU and memory allocation (see our guidance on choosing resoures) that is most appropriate for your work. Keep in mind that there are different costs associated with each configuration. Starting a launcher for the first time can take 5-10 minutes.

image

Once your session has been provisioned, click on "Connect" to open the software in your browser.

image

Starting Sessions

Once a launcher is provisioned, it will appear as one of your sessions at the top of your project space Workbench when you log into the RCP. Click on the "Start" button to re-start the session. It will take a few minutes to load, and then the "Connect" button will appear. Click on the "Connect" button to open the software in your browser.

image

image

Stopping Sessions

If you have finished with your session, but plan to return to it, please ensure you have stopped it. The Stop option will appear after clicking on the hamburger menu to the right of the Connect button. Keeping the session running, even when you’re not actively using it, will incur charges, whereas a stopped session only incurs minor storage costs.

image

Terminating Launchers

If you anticipate no longer needing a launcher, please terminate it. The Terminate option will appear after clicking on the hamburger menu to the right of the Start button of a stopped session. A terminated launcher will incur no costs. Keep in mind that once you terminate a launcher, downloaded packages (e.g., R packages) will be removed and unavailable for future use (i.e., if you restart a launcher you will need to recreate your environment).

image

Guidance on Choosing Resources

Each launcher allows you to select a combination of CPUs/GPUs, and memory (RAM). Choosing the right configuration helps you balance performance and cost. Use the guidance below to choose settings that match your workload and budget.

Memory (RAM)

⚠ WARNING
Please note that currently when a job on RCP runs out of memory, it stops but does not produce an error message to the user. RCS/IT is aware of this issue and working on improvements.

A helpful rule of thumb is:

Request RAM that is approximately 4–10× the size of your dataset.

Why?

  • Text files like CSVs expand significantly when loaded into memory — sometimes up to 10× their on‑disk size.
  • Binary formats (RDS, Parquet, Feather, Arrow, .fst) expand only 1.2–1.5×.
  • Best practices recommend avoiding both under‑ and over‑provisioning.

Practical Tips

  • If your workflow runs successfully on your local computer, start with a similar RAM or less (for example, if your laptop has 8GB of RAM, try asking for 8GB).
  • For CSVs, assume 10× expansion when loaded into memory.
  • For binary formats, assume 1.2–1.5× expansion.
  • Convert CSVs to binary formats to reduce RAM usage and improve load times.
  • If you know your past memory usage, request that amount plus 20% overhead to avoid crashes.

CPUs and GPUs

CPUs

A CPU handles general-purpose computation. Most tools such as R, Python, Stata, Spyder, and MATLAB use one CPU by default unless explicitly parallelized.

GPUs

A GPU accelerates massively parallel operations and is useful only when your code uses GPU-enabled libraries, such as:

  • TensorFlow, PyTorch, or JAX
  • CUDA-enabled simulation frameworks
  • Transformer-based NLP models
  • Large-scale deep learning workloads

If your code is not GPU-enabled, a GPU instance will be more expensive and often slower than a CPU-only configuration.

Quick Tips for Choosing CPUs

  • Use 1 CPU for almost all interactive work, including R, Python, Stata, Spyder, and Jupyter notebooks.
  • Choose 2+ CPUs only when your code is explicitly parallel-aware, such as:
  • Python: multiprocessing, joblib, ray, PyTorch DataLoader
  • R: future, parallel, data.table multithreading
  • MATLAB: Parallel Computing Toolbox
  • Stata: limited multiprocessor support via set processors
  • More CPUs do not improve performance unless your code is parallelized.
  • Large CPU allocations should be used for short intensive jobs, not long-running sessions sitting idle.

Quick Tips for Choosing GPUs

Use a GPU if:

  • You are training neural networks
  • You are working with transformer-based NLP models
  • You are performing large-scale tensor operations
  • Your code uses CUDA or GPU-enabled libraries

Do not use a GPU if:

  • You are performing regression, modeling, or typical statistical tasks
  • You are cleaning data or visualizing
  • You are working in R or Stata without GPU packages
  • You are not intentionally using deep learning libraries

Cost Impact Legend

⚠ WARNING
If you anticipate installing many packages in an RStudio launcher, please note that this can take a long time and stall when using the smallest configuration. Given this, if you will be installing many packages in RStudio, you may consider using a configuration one level larger than what is recommended below. RCS/IT is aware of this issue and working on improvements.
  • 💲 = Low cost
  • 💲💲 = Moderate cost
  • 💲💲💲 = High cost
  • 💲💲💲💲 = Very high cost
Use Case Recommended Config Example Launcher Cost Impact Notes
Light interactive work (R/Python/Stata) 1–2 CPUs, 4–8 GB RAM Medium General Purpose (2 CPU / 4 GB) 💲 Good for small datasets and exploration.
Medium analysis (2–10 GB CSV) 1 CPU, 16–32 GB RAM Memory-Optimized Large (2 CPU / 16 GB) 💲💲 CSV expands significantly; convert to binary formats.
Heavy analysis (10–30 GB CSV) 1–2 CPUs, 32–64 GB RAM Memory-Optimized XL (4 CPU / 32 GB) 💲💲💲 Suitable for large merges and joins.
Large binary datasets (10–50 GB) 1–2 CPUs, 16–32 GB RAM Extra-Large General Purpose (8 CPU / 32 GB) 💲💲💲 Binary formats reduce RAM needs.
Stata large merges 1–2 CPUs, 16–32 GB RAM Stata Large General Purpose (4 CPU / 16 GB) 💲💲 Stata benefits from modest multiprocessor usage.
Parallel simulation 2–8 CPUs, 8–16 GB RAM Compute-Optimized Large (8 CPU / 16 GB) 💲💲💲 Only useful if your code is parallelized.
Classical ML (sklearn/caret) 2–4 CPUs, 8–16 GB RAM Large General Purpose (2 CPU / 8 GB) 💲💲 Most classical ML does not need GPUs.
Deep learning training 4 CPU, 16–32 GB RAM, GPU Accelerated ML Instance (4 CPU / 16 GB GPU) 💲💲💲💲 Recommended for neural networks.
Heavy NLP (transformers) 4–8 CPU, 16–64 GB RAM, GPU ml.g4dn.xlarge or similar 💲💲💲💲 GPU strongly recommended.
Enterprise-scale DL 48 CPU, 192 GB RAM, GPU g6extralarge GPU instance 💲💲💲💲 Only for extremely large workloads.