Linux Scaling Benchmarks With The AMD Threadripper 2990WX In Various Workloads

[phoronix]

Phoronix: Linux Scaling Benchmarks With The AMD Threadripper 2990WX In Various Workloads

While yesterday were the benchmarks showing how Linux games struggle to scale past a few CPU cores/threads, in this article is a look at the scaling performance of various applications/workloads under Linux up to 64 threads using the AMD Threadripper 2990WX. Here's a look at how the Linux performance changes in a variety of applications from one to sixty-four threads with this new HEDT processor.

http://www.phoronix.com/vr.php?view=26722

[chuckula]

You need to update your graph software so that numbers don't have to be displayed inside the bar since many of those are unreadable.
Also, semi-log is your friend in these situations. Especially if it's base-2 so it aligns with the core counts.

[Michael]

You need to update your graph software so that numbers don't have to be displayed inside the bar since many of those are unreadable.
Also, semi-log is your friend in these situations. Especially if it's base-2 so it aligns with the core counts.

I don't often use the vertical bar graphs, the horizontal bar graphs handle that case fine, just haven't bothered porting it over to vertical graph for lack of time.

[schmidtbag]

These charts do a great job exemplifying Amdahl's Law. It seems for most modern heavily-multithreaded applications, 32 threads is the most you really need until you can't efficiently take advantage of the whole CPU.

[Weasel]

These charts do a great job exemplifying Amdahl's Law. It seems for most modern heavily-multithreaded applications, 32 threads is the most you really need until you can't efficiently take advantage of the whole CPU.

It's because it has only 32 physical cores.

[Spring]

The graphs would not be very informative with log scaling either. The valuable information which graphs should express is not the absolute value but relations between adjacent values.

[AnonymousCoward]

This seems pretty expected, the Windows comparison will be interesting.

BTW, pet peeve, the word "etc" is really useless, either there is more to say, then say it, or just leave it out.

[DavidKL]

Article contains useful info as always. Thanks.

[xorbe]

Can't see the numbers on the shorter bars.

[thelongdivider]

Very nice as it is good to see how things scale in various scenarios. More than that, this compares (at 64 vs 32) the effect of SMT vs real execution cores. Most impressive was in the kernel compilation, where SMT only shaved a single second off the compile time (a ~4% reduction).

Thanks for the tests!