Hi,
I have a friend who is looking to run a few simulations he has implemented in python and needs around 256GB of ram. He is estimating it will take a couple of hours, but he is studying economics so take that with a grain of salt 🤣
For this instance, I recommended GCP, but I felt a bit dirty doing that. So, I was wondering if any of you have a buttload of memory he can burrow? Generally, would you lend your RAM for a short amount of time to a stranger over the internet? (assuming internet acccess is limited to a signle ssh port, other necessary safeguards are in place)
Needing that much RAM is usually a red flag that the algo is not optimized.
Nope. Some algorithms are fastest when a whole data set is held into memory. You could design it to page data in from disk as needed, but it would be slower.
OpenTripPlanner as an example will hold the entire road network of the US in memory for example for fast driving directions, and it uses the amount of RAM in that ballpark.
Sure, that is why I said usually. The fact that 2 people replied with the same OpenStreetMap data set is kinda proving my point.
Also, do you need the entire US road system in memory if you are going somewhere 10 minutes away? Seems inefficient, but I am not an expert here. I guess it is one giant graph, if you slice it up, suddenly there are a bunch of loose ends that break the navigation.
*looking at the 14TB cluster I had running for 18 hours
Yep, nobody would ever need that much memory
Wow, yea I think you win that contest lol.
To be honest, it was a very paralell process. I could do a fraction of the compute, needing a fraction of the RAM, but taking a shit ton more time.
Also, theres no perfect machine for this use. I can have 3.5 times more RAM than needed, or start swapping and waste time.
Researchers always make some of the worst coders unfortunately.
Scientists, pair up with an engineer to implement your code. You’ll thank yourself later.
True, but there are also some legitimate applications for 100s of gigabytes of RAM. I’ve been working on a thing for processing historical OpenStreetMap data and it is quite a few orders of magnitude faster to fill the database by loading the 300GiB or so of point data into memory, sorting it in memory, and then partitioning and compressing it into pre-sorted table files which RocksDB can ingest directly without additional processing. I had to get 24x16GiB of RAM in order to do that, though.
Yea, that makes sense. You could sort it in chunks, but it would probably be slower. If you are regularly doing that and can afford the ram go for it. Otherwise maybe extract the bits that need to be sorted and zip them back up later.