Benchmarking The Current Spectre + Meltdown Performance Overhead For 10 GbE Networking
While running the Windows vs. BSD vs. Linux 10GbE network benchmarks among other recent 10GbE Linux network performance figures, the test request came in from a premium patron to look at the current 10GbE network performance hit as a result of the default Spectre+Meltdown mitigations.
I ran some tests on an Ubuntu installation with the Tyan dual Xeon server used in the recent comparisons. With that Ubuntu 18.10 install, the Linux 5.0 Git kernel was running for the most recent kernel experience. Of course, on particularly older kernels the impact may vary with the Spectre/Meltdown mitigations continuing to be refined among other kernel optimizations to help offset the induced overhead.
Long story short, here's a look at some of the changes when comparing Linux 5.0 booted out-of-the-box and then again when using the relevant boot-time switches to disable the mitigations that can be diverted at run-time.
Microsoft's Ethr network benchmark was showing some performance hits. In some workloads we've found for I/O and now as well for networking to be slightly faster actually with these mitigations due to the CPUs being kept busier and in a slightly higher performance state.
At least from the common network benchmarks run, for the most part the 10GbE network performance with Linux 5.0 wasn't all that impacted by the current Spectre/Meltdown mitigations. The main test case adversely affected was Ethr's latency tests while the throughput tests were largely unaffected.
I ran some tests on an Ubuntu installation with the Tyan dual Xeon server used in the recent comparisons. With that Ubuntu 18.10 install, the Linux 5.0 Git kernel was running for the most recent kernel experience. Of course, on particularly older kernels the impact may vary with the Spectre/Meltdown mitigations continuing to be refined among other kernel optimizations to help offset the induced overhead.
Long story short, here's a look at some of the changes when comparing Linux 5.0 booted out-of-the-box and then again when using the relevant boot-time switches to disable the mitigations that can be diverted at run-time.
Microsoft's Ethr network benchmark was showing some performance hits. In some workloads we've found for I/O and now as well for networking to be slightly faster actually with these mitigations due to the CPUs being kept busier and in a slightly higher performance state.
At least from the common network benchmarks run, for the most part the 10GbE network performance with Linux 5.0 wasn't all that impacted by the current Spectre/Meltdown mitigations. The main test case adversely affected was Ethr's latency tests while the throughput tests were largely unaffected.
4 Comments