A Look At The MDS Cost On Xeon, EPYC & Xeon Total Impact Of Affected CPU Vulnerabilities

Next is looking at the performance impact of mitigations=off / mitigations=auto (all relevant default/out-of-the-box CPU mitigations applied currently by the Linux kernel respectively for Intel and AMD) / mitigations=auto,nosmt (disabling SMT / Hyper Threading where deemed unsafe, Cascadelake didn't see HT disabled in this configuration and obviously not relevant for AMD). This round of tests was done on a dual socket AMD EPYC 7601 server, dual socket Xeon Gold 6138 server, and dual socket Xeon Platinum 8280 Cascadelake server for looking at the overall impact of these mitigations on each of the different boxes.

There was varying RAM between the systems due to different memory channels and other factors, but each server did see all memory channels occupied with RAM at each of the processors rated frequencies. All three of these servers were tested with Ubuntu 19.04 running the patched Linux 5.0 kernel with the latest microcodes for each processor under test.

PostMark mostly exercises fsync and for this testing the EPYC server saw a 3.5% hit from its relevant Spectre mitigations while the Xeon Platinum 8280 saw a 4% performance hit from its relevant mitigations. The Xeon Gold 6138 server was most impacted with a 18% hit to the performance due to not seeing any hardware mitigations unlike the brand new Cascadelake processors.

The Sockperf throughput performance was actually least affected by the Cascadelake server with losing just 3% performance while the current-generation EPYC 2P server saw a 12% performance drop from Spectre but at least not as bad as the Xeon Gold 6138 server losing 19% performance in the Sockperf benchmark.

The latency was obviously the worst impacted by these mitigations as well for the 1st Gen Xeon Scalable.

Server Address: localhost - Protocol: HTTP - Test: Bandwidth - Threads: 1 Server Address: localhost - Protocol: TCP - Test: Bandwidth - Threads: 64