Pandas Dataframes on your GPU w/ CuDF

Key Takeaways

I of course like many packages in libraries in Python but probably my favorite package of all time is pandas I love it because it's just plain super easy to use and makes sense for a lot of my workflows in data science that said I also like machine learning and data sets can often get quite large for example we'll play with the data set today from kaggle of UK property prices with over 28 million rows of data using vanilla pandas for this despite how awesome it really is as a library and how efficient it is out of the box arguably uh it would and will be very slow in the past I've simply employed the skill of patience for this problem but today I'd like to show you the pandas accelerator from nvidia's Rapids Arsenal using KF which is a GPU accelerated data frame Library historically this library has been a bit of a challenge to deploy because if you wanted to use CDF you needed to completely adopt qdf and make sure every operation that you performed in your program was function within CF and given the scope of operations even in a simple project this made CF almost unusable in practice no matter how fast it ran for certain specific commands and then even if you did have full and perfect coverage you'd still have to refactor all of your code which no Dev really wants to do and then you'd probably have to maintain separate branches and all that nobody wants that instead what Nvidia has done is introduced a way to deploy C with a mere module flag when running the script or through an extension load in a notebook and that's it there's no other change necessary with a slight asterisk of maybe sometimes being explicit with data types uh otherwise you might run into some problems and this came up in my testing and I'll share with you all what I ran across U but at least in my testing so far this is basically fair to say a dropin functionality replacement that you should could absolutely be using right now today and from now on with your Panda data frames this also includes even if you just so happen to be using a library that also uses pandas itself because it will also accelerate that too in order to install it's just a pip install cf- q11 or q12 for the Cuda version that you're running so let's see just how much performance we can get with a mere flag I'm going to run the CF pandas accelerator on the left and then regular pandas on the right we could use the kind of time it magic and all of that but this does weird things with variables and is really only useful for getting a really solid concrete sort of average time um so my test here isn't going to be the most perfect of timing tests but it's a real comparison nonetheless and the kind of the difference between CDF and regular pandas is on the orders of like magnitudes right so so um we're not going to have um an issue with a ten of a second here and there there so anyways uh just know that this is basically it's just one pass through it's not the most perfect comparison but it's good enough because the difference in performance here is massive so the first sell in this notebook is just loading the accelerator for CF and then I just have like a comment for the other one which is our regular pandas between the two notebook files this is actually the only difference these are otherwise identical notebooks uh so next what we're going to do is we're going to load the data into a data frame note that I'm explicitly setting the data type for price as int64 and this is what I was talking about earlier the reason for this is in the data set itself in that CSV file it has these prices in in quotes like a string pandas with regular old python actually still magically handles for this and allows us to perform math operations on strings as if they were ins um or even floats and this is just I don't know some magic going on in the back end of of pandas to kind of handle for this when people do this um but but if you are having a problem with like in Infinity numbers and stuff like that you'll want to go ahead and look at your data set and see if it's doing this cuz in my opinion this isn't really I'm not sure this is something I would expect anybody to handle for even pandas but just keep in mind that this could be a thing in your data set that panas is just somehow magically handling for but anyways you're going to want to either make sure your data set doesn't do this or you can explicitly set data types to handle for it just in case it does as you can see even just loading the data frame into memory sees a mass M boost of a mirror only like 2.9 seconds to load with qdf compared to almost a minute with regular pandas the first time I ran through with CF I actually found that both took quite a bit of time to load after some debugging I realized that it was actually me loading from my Nas rather than from like a local storage on that machine so network storage versus local storage so just in case you're as dense as I am uh make sure your data is actually local so in this case it's on an nvme so it's much much faster to load instead of of loading over Network once it's in memory though it's kind of irrelevant where the data set started from so it's all kind of a wash from there okay so from this point we can do a quick operation just to make sure things are working as intended and we'll get the length of the data frame which looks good and is as expected so then what we can do is grab all of the unique towns and cities interestingly enough there is a slight Delta here in favor of regular CPU pandas the reason here is likely this is a CPU only oper ation so then you're paying for some data transfer and overhead going CPU back to GPU and I think that's the difference that said you're about to see why this difference is insignificant and irrelevant but you might also wonder what if you did a larger amount of these unique operations I don't think I've ever done a unique call in a loop but what if you did since I already had to redo this video once I went ahead and just did a quick test for this particular operation I looped through just over all of the columns and grabbed uniques from each of them and it's still just a CPU only operation but you can clearly see uh it's not like CPU to GPU transfer is actually causing a 2X slowdown if you were to need to scale out a unique operation for any reason um and actually it even looks like the kdf version was slightly faster for this run that's probably just by chance if like I I was saying before if we use the time it module for like little tiny differences like this we might find that either they basically identical or maybe sometimes the the CPU is faster or whatever but anyway I just wanted to further poke into this and make sure that the overhead was actually as irrelevant as I was pretty sure it was and in this case it definitely is so now what we're going to do is get the average home price for each unique town or city this operation takes qdf 10 and 1 12 seconds for regular pandas it takes 19 minutes uh 10 seconds versus 19 minutes is absolutely insane and again this is a zero effort switch if you already have an Nvidia GPU you've already got Cuda all that it's it you just install KF you call the flag and boom you're done the same code also can run on the GPU if it's available or it could be CPU only so you don't even need two separate code bases continuing on what we're going to do is we'll slice out just the year 2022 from this data frame and get just pricing information for that year surprisingly to me this operation was also orders of magnitude faster with KF than without this is surprising because these dates were not even read in as daytime objects or anything fancy at all they're actually simply strings and the actual operation that per we're performing here to do that slice is actually just a string dot starts withth so this means there's actually some fancy kouf GPU method for string operations which is particularly impressive and something that I did not at least personally expect to see here and that brings us to the profiler this lets us see exactly what's happening either closer to the qdf backend showing us the operations running on cdf's side of things and what device it's running on GPU or CPU and then how long that's taking alternatively we can also use the line profiler to show us sort of line by line what's running on the CPU or GPU and how long that's taking so here we can see both the line profile and just regular profile on that data frame slice line profile shows us that actually that line of code is indeed done entirely on the GPU we can then use the profiler to see what KF is actually doing and we can see that there's a get item method and then that we do indeed have string Methods which includes some fancy starts withth functionality allowing this to run entirely on the GPU despite the fact that it's just string operations which is particularly awesome next maybe we want to calculate the upper and lower 20% of home prices to discover areas with the most variance so here's a slightly longer bit of code we can again see that CDF is considerably faster here by about five times and we can also do a line profile here to see what's going on finally from here after pulling all this kind of data we might do something like some pure python code to grab ratios for high and low as well as maybe get an actual range and then kind of display maybe where are the biggest variants and stuff like that so then finally just a couple of quick common commands like finding the max for a column and a sort to just show off some more operations but I really encourage you to install CF and just use it on your own projects and stuff like that just to see how much faster it is and see what else might even just surprise you cuz like I said I I had seen some of the demos but then there were things that I just did not expect to be the case and that were definitely the case and and that was pretty cool to see so anyway uh you can try the pandas uh accelerator yourself yourself via CDF by installing with Pip install cf- q11 or Q2 also of note you can run scripts too so obviously not just notebooks and the way that you can do this is with just the kf. pandas module flag when you go to run that actual script and you'll get all the same performance speed UPS so that's all I have to show for now this pandas accelerator from KF is freaking awesome I love how darn simple the implementation of it really is and I look forward to seeing how this like rapid stuff just evolves over time this is been many years in the making and this is oh such a cool update to see so if you have an Nvidia GPU and you find yourself using pandas definitely try it out let me know what you think uh if you find anything in particularly interesting or if you're having a hard time with something leave a comment below um I really was just pretty surprised at just how fast everything actually is how well everything just simply works if there isn't something it just falls back to the CPU but like I said there was more coverage for GPU operations than I even expected like the string Methods and that was just super super cool to see I'll leave uh pip install commands and an example of running the script via the terminal in the description there's also an official collab notebook example from the Rapids team that I will put in there as well and then finally there was an AI data science Summit from Nvidia very recently that also showcased this uh Panda's accelerator which you can also watch I'll put a link to that in the description as well anyway I hope you all have enjoyed and learned a cool new tool I'm definitely going to be using this personally from now on and I will see you all in a another video