Slashdot Mirror
Linux Kernel Benchmarking: 2.4 vs. 2.6-test
frooyo pastes from kerneltrap: "Cliff White recently posted some re-AIM multiuser benchmark results comparing the stable 2.4.23-pre5 kernel against the 2.6.0-test5 and 2.6.0-test5-mm4 development kernels. In his conclusion he makes reference to earlier scheduler tests posted by Mark Wong saying, "Short summary: we mostly rock.""
It's only significantly faster if you have 8 processors.
Whereas it is 7% slower if you have one processor.
I suppose they'll have uniprocessor version which runs faster? Lots of people have uniprocessor pcs.
Hyperthreading doesn't really count.
I'd love to see Linux being the best OS for multiple CPU scaling.
You do need a scalable OS to suport lots of processors, of course, but you also need hardware that scales too (clustering doesn't count). Example - SGI is using Linux with NUMAflex on the Altixes to cluster 64-processor system images, but that kind of hardware isn't commodity in any way, and isn't going to be anytime soon.
Anyway, Linux doesn't scale THAT well...as of 9/2000, SGI was using IRIX for a 1024-processor single-system-image supercomputer; I've heard they can go to 2048 now, but I don't have anything to back that up. Dunno about Solaris, but I imagine it's pretty scalable as well.
Facts do not cease to exist because they are ignored. - Aldous Huxley
I've got nothing against Linux improving at SMP in essence, but there is something very bad going on here it seems to me. Notice that while the new kernel 'kicks ass' on SMP systems, on uniprocessor systems the 2.4 kernel is the one kicking ass. Anyone benchmarked 2.4 against some of the pre-SMP kernels on a uniprocessor machine?
Face it, the vast majority of users are uniprocessor, and kernel performance is more of an issue on lower-end machines. Improving performance on big multiprocessor boxes is fine by itself, but not when it harms uniprocessor performance. I'm not a kernel hacker, but I've read many people that this would not happen, that the SMP code would not hurt performance on a uniprocessor machine when the kernel is compiled without it, but that's obviously not turning out to be the case. Anecdotal evidence, at least, suggests that this performance degradation has actually been going on for quite some time, at least back to when SMP code first started being added.
I'm not sure what all the factors here are, so naturally I'm not going to tell you the solution, but it certainly looks like a potential problem that should be discussed. Hopefully someone with more specifics than I have can chime in...
=-=-=-=-=-=-=-=-=-=-=-=-=-=-
Friends don't let friends enable ecmascript.
linux-2.6.0-test5 - 992.06 - Uni
linux-2.6.0-test5 - 1017.43 - Dual
linux-2.6.0-test5 - 5406.68 - Quad
Does this mean that you only gain 3.49% when adding a 2nd processor? Obviously I don't expect things to scale linear but 3%!? Am I missing something here? And then 81.65% for quad? I'm not trolling, I'm looking for someone to explain what I'm missing.
You are correct! The scheduler reacts different to different work loads. This is why the kernel developers try hard to test their changes under a number of different workloads. To top it off with, they attempt to target the benchmarks which behave like real-world work loads rather than contrived and unrealistic workloads. That's not to say that they don't test those too, however, they clearly direct more attention at real-world workloads and corrosponding result sets.
The 2.6x series kernels will be a big step up for just about everyone that seriously uses their computer. Significant realiability improvements as well as faster thoughput on disks, much, much higher scalability for SMP (hyperthreading and numa and even highly loaded uni-systems) systems, and much lower latencies, all at the same time. Granted, there are still some tests which may not be a win-win all the way around, however, almost everything in general is an improvement with hardly any detracters.
So, saying, "we mostly rock", really is a true statement!
That statement simply is not true. Granted, you can always find some corner case where the workload is going to be slower between releases (2.x or 2.6), however, as a rule of thumb, 2.6 should still be a huge improvement for even uniprocessor users. Best yet, many, many parameters of the kernel and scheduler are tunable, so, you can always adopt the kernel to work best for your specific workload needs.
While it's true that they are working hard to significantly improve Linux for the server room, by far, they have never lost site of the uniprocessor user. Remember, there is nothing wrong with tuning the kernel for your uniprocessor needs, and specific workloads. They just can't do that when they are benchmarking because it would skew the results, invalidating them. They are not only trying to measure how their improvements effect the overall system, but, what makes for sane initial defaults, which are reflective of a general purpose and broad workload. If you understand what you are doing, there is not a reason to believe that you can't greatly improve things for your specific uses and workloads. It's important to keep all of these in mind when talking about these benchmarks. Furthermore, you should fully expect your favorite distro to come with tuning presents which reflect a targeted workload (file/print server, workstation, database, web server, etc.).
Keep in mind that the benchmark you looked at represents one category of many different types of workloads. So, for that specific workload, it may of been slower, however, that workload my not represent anything you do with your computer. Remember, other types of workloads are significantly faster. One last note, remember, performance is the classic trade off with lower latencies. It trades responsiveness for raw throughput. If, on a uniprocessor workstation, you only see a -7% drop in performance and latency is greatly reduced, chances are, not only will you never notice the loss in performance, but you'll be praising it for how well it works with your mouse, monitor and keyboard (if feels better and makes you a happier user).
Just some food for thought.