Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Performance of Stata on Linux and Windows

    Hello, we provide Stata/MP on our RedHat Linux computing cluster. I have had what seems a rather bizarre conversation with a research group who are thinking of using the cluster do to use Stata. They seemed to intimate that Stata was faster on Windows than on Linux and hence could we provide some Windows resources in the cluster. Generally speak, I would assume the converse to be true, however is there any evidence to suggest that Stata/MP is slower on Linux? I have seen some posts, dating back to 2005, which appeared to suggest that Stata/MP was faster on Windows, but I'm struggling to believe could be true today. Does anyone have any thoughts in this request, please.

  • #2
    You can ask [email protected]

    Comment


    • #3
      Can you link to the "posts dating back to 2005" here? I guess I could imagine that if the Stata devs work primarily on Windows, Stata might be better optimized for the Windows OS. Of course, that doesn't mean windows is a better solution in your cluster environment just that it might be better all else equal. It also depends on the use case. For example, I'm pretty sure Windows will not let a single application consume more than 2 gigs of memory at a time - at least that was a constraint on my C#/.NET applications on windows professional back in 2019. You can relax that memory constraint in linux. So if they have a problem with a lot of data they may be able to consume more memory and avoid needing to use virtual memory on the disk with linux.

      You could always test this. Maybe create a relatively difficult test case like the merging of two very large datasets. Try to keep the hardware constant either with two virtual machines on the same computer or with a fresh install of the operating system for each test. Then time the process and see which execution environment is faster.

      Comment


      • #4
        Thank you for your replies to my question. I agree that the first step here is to get a representative do file and data from the reseachers so that we can do some benchmarking. The post that I referred to is here -- https://www.stata.com/statalist/arch.../msg00468.html. The timing (19 minutes vs 12 minutes) in the post are, well, trivial and don't fill me with any sense of confidence! In other words, the first step has got to be to run the benchmarks and see how they look. Given that our cluster environment is heavily based on Linux (Slurm scheduler, etc) then we have to be sure before we start to add an Windows machines.

        Comment


        • #5
          OMG.

          The link in post #4 takes us to a single anecdote from 2005 comparing Stata 9 performance on 64-bit Linux on one computer to Stata 9 performance on 32-bit Windows on a different computer.

          17 years later we are 8 major releases later with Stata, over 10 generations of Moore's Law later for the hardware, and who knows how much further along on the operating systems.

          In other words, the first step has got to be to run the benchmarks and see how they look.
          I disagree. That's seems likely to just provide updated anectdata.

          From this linked page of Stata/MP marketing you can follow the link to the Stata/MP Performance Report which at this writing was most recently updated a year ago. While the thrust of this report is comparing Stata/MP performance across a range of numbers of processors, the document tells us that "Stata/MP performance was tested on many computers under MS Windows, Mac, and Linux operating systems." So the recommendation in post #2 was not lightly made. Perhaps Stata Technical Services can elaborate to you on their extensive experiences with different platforms.



          Comment


          • #6
            Perhaps Stata Technical Services can elaborate to you on their extensive experiences with different platforms.
            I have to agree, if for no other reason than because a statement from Stata Technical Services will likely be most compelling to your potential client.

            Comment


            • #7
              I disagree. That's seems likely to just provide updated anectdata.
              Well, point taken, but as any lawyer or qualitative researcher will tell you, anecdotes are a kind of evidence, and can be very useful if one remains mindful of their limitations. I still see value in benchmarking OP. Then you might be able to say to your client "We did some independent benchmarking and confirmed the statement from Stata."

              Comment

              Working...
              X