I recently became interested in using BigDFT, but I still have a few questions.
1. What is the computational performance gain of BigDFT compared to more traditional DFT packages (such as VASP, Quantum ESPRESSO, or GPAW)?
I’ve read that BigDFT uses wavelets and is said to be efficient for large systems. But in practice, does the simulation time really decrease significantly? In which cases is the computational gain most noticeable?
2. I’ve seen some runs with around 2000 atoms using 6000 cores that completed in just a few hours. Are these “cores” usually CPU threads or GPU units in the context of BigDFT?
I’m asking because I’d like to understand what this means in terms of more accessible hardware. For example, if I have an RTX 3060 GPU, would it be possible to perform similar simulations, or are supercomputers required?
3. Is there any practical equivalence between a consumer GPU (such as an RTX 3060 or 3090) and a certain number of CPU cores for running BigDFT?
ust to get a rough idea: how many CPU cores (on average) would match the performance of a GPU like the RTX 3060 in BigDFT simulations? Is there an approximate comparison?
4. Does BigDFT scale well for large simulations? For example, if I double the number of atoms, does the simulation time also double? Or does it scale more efficiently?
I’m trying to understand whether the method is truly linear for large systems, as sometimes mentioned. Does the scaling behavior change depending on the type of material (e.g., organic vs inorganic)?
5. Do I need a specific compilation to run BigDFT with GPU acceleration? Or does it work directly with CUDA, OpenCL, etc.?
Hello Lucas Faria da Silva
I started to use BigDFT to compare with Quantum ESPRESSO. Until now, I only tried small structures, but I could see some differences already. For systems like BCC Iron and Perovskite CsPbBr3, which are small systems, cubic symmetric, they are very fast to in Quantum ESPRESSO. In BigDFT, I noticed they are slower, using the Input I made. I still didn't explore parallelization in BigDFT, but I know it can do. The memory used is also different. For this CsPbBr3 you can see in the output that it required 5 GB, compared with only 250 MB in QE. I don't know if I could make BigDFT use less memory, because 5 GB for this perovskite is too much. I am going step by step for larger systems, because I don't have access to a cluster, so I will try until I can.
I would like to keep in contact to explore more about this code.
Feel free to reply
Best regards,
Dr. Ricardo Tadeu.
Hello Ricardo Tadeu Maia! Thank you very much for being here!
Indeed, I agree with you. 5 GB for a system like that does seem quite high, even though I'm not deeply familiar with your specific perovskite structure. While I have the chance, I would like to clarify a question: how does the parallelization work in Quantum ESPRESSO? Does the code typically scale well with an increasing number of processors, or is there a point beyond which the performance gains become marginal?
Also, I’d like to ask if you’ve had the opportunity to perform molecular dynamics simulations using Quantum ESPRESSO. I’d really appreciate hearing a bit about.
All the best.
Hello Lucas Faria da Silva
I am no expert in Parallelization, but I know how to change it in Quantum ESPRESSO. You can control the number of nucleus and number of threads QE can use. I use QE on WSL or directly on Windows in my PC, so I definitely don't push it hard. Usually I go for 1 nucleus and 8 threads. I noticed that for small cells like diamond, silicon, beryllium, 8 threads can be slower than use only 2 threads. But for larger systems, 8 threads seem better.
Check this question I added here in ResearchGate:
https://www.researchgate.net/post/How_to_set_parallel_calculations_in_Quantum_Espresso_using_Ubuntu
For molecular dynamics, I tried for some systems, calculations ended normally, but kind of I don't know what I am looking for with MD. I just know how to execute, but still don't know how to extract import informations with it. Let's keep in contact because I know its a valuable tool, I just need to understand what I can do with it.
Hi Ricardo Tadeu Maia! It would be a pleasure to stay in touch and possibly ask you more questions in the future! I tried reaching out to you on LinkedIn, but I haven’t received a response. Would you mind sharing your email with me?
- Badges
- Science topic
- Similar topics
- Social Science
- Media
- Journalism