6/22    I'm doing some performance tuing of a multithreaded application on a
        dual-Xeon with hyperthreading.  Windows reports 4 cores and counts
        1 core @ 100% as 25% CPU usage.  I realize this is a fuzzy question,
        but what's a reasonable real-world limit for CPU usage in this
        situation, since there aren't really 4 cores?  This app has a mix of
        MMX computation, unaccelerated floating-point math, bus I/O, and a
        lot of pointer dereferencing / irregular memory access patterns.
        \_ Turn off hyperthreading.
           \_ Since I currently get around 60% reported CPU usage, it would
              seem hyperthreading is helping performance about 20%, right?
              \_ The general wisdom is that hyperthreading increases GUI
                 responsiveness but decreases overall performance by 0-5%.
                 I'm sure there's an example where hyperthreading increases
                 overall performance, but I can't think of one.
              \_ Run your app with it on and off. Compare the run times.
                 \_ Just tried that.  Hyperthreading on:  200 seconds
                                      Hyperthreading off: 220 seconds
        \_ I'm doing the same with a simulation app.  It hits the memory bus
           and CPU most.  HT actually slowed our system down considerably.  A
           dual xeon gave us about 2x speed improvement.  The dual opteron did
           the same.  We're now checking a dual opteron with each package
           having dual core.  I say this as a former Intel employee who likes
           Intel--HT sucks for anything other than casual use.  Hard. -emarkp
           \_ I agree.  If yer doing simulations or anything else hardcore,
              turn off HT.  If yer doing Word and net browsing, by all means
              turn it on because it's a more pleasant user experience.
              Or wait until the Athlon dual-cores drop from $630.
              Or don't wait and get an Intel dual-core for $250 if your room is
              too cold.
        \_ OK, so aside from dissing hyperthreading, nobody has given actual
           information on what is a realistic maximum load for a 2-Xeon
           hyperthreading system. -OP
           \_ I think you should rephrase your question. What are you
              asking? What parameter(s) do you wish to change that are
              under your control? The load should be 2 at most, since
              there are 2 CPUs. Play with it and see what happens when you
              run more processes.
        \_ Do you have two dual-core CPUs with hyperthreading enabled?
           That implies eight logical CPUs.
           If you disabled hyperthreading, then you have four logical CPUs
           and four real cores.  You should get theoretical 100% usage on
           each core.  The simplest way to do this is to write a C program
           with a while (1) loop and run it four times.
           We have an old 700 MHz 4-CPU machine at work, and I can get about
           that running Linux with top.
           For one or more real-world apps that each uses multiple threads,
           it really depends on the threading implementation in each program,
           followed by how good the kernel is at handing out tasks to 4+ CPUs.
           I can easily imagine a two-thread program that uses 100% CPU on
           one CPU and 25% CPU on a second.
           \_ I have two dual-core dual processor machines with hyperthreading
              enabled ......
              \_ So each machine has one dual-core with hyperthreading?
                 Yeah, then that means four logical CPUs, two real cores,
                 per computer.
                 You can only expect a maximum of 100% CPU usage over the
                 two logical CPUs that sit on one physical core.
                 Anyways, if you run four while (1) C programs you should
                 get 200% total over the four CPUs as reported by top.
                 Same thing for Windoze if you have four CPU graphs.