Hector Martin

Ha, Intel were *so proud* of their high throughput RDRAND, and now it turns out they leak the values all over the other cores and the microcode patch to fix it... has a 97% performance hit. As in you get ~1/30th of the performance you used to. Whoops! https://www.phoronix.com/scan.php?page=news_item&px=RdRand-3-Percent …

As @sebsued points out, it may well become a DoS vector itself, because if I understand the mitigation right, RDRAND ops are now *serialized against the entire chip memory bus*. They block all other cores from doing any reads and writes that miss cache. https://twitter.com/sebsued/status/1270639456312311809?s=19 …

I just did a test. One thread of RDRAND on an unpatched Xeon E-2144G (8 threads) drops 7-threaded sysbench performance by 10%. The same with the patch, 55%. So with the patch, a rogue core can now halve your memory performance. But it gets better.

Running 8 threads (on an 8-thread E-2144G) of RDRAND in the background, memory throughput drops by *81%* with the patch vs unpatched (91% vs idle). So if someone can spawn a bunch of RDRAND threads, they can drop your memory perf to 19% of what it would be, 9% of nominal.

I didn't find the time to run this test for the paper, so thank you so much for actually doing it! It was clear it'd be bad given that they put the effort in to add an MSR to disable the mitigations again. It's about the level of performance hit I expected, but.. ouch. Very ouch.