On 11/02/2018 09:42, Wanpeng Li wrote: > Cc chandu, > 2018-01-08 22:53 GMT+08:00 Paolo Bonzini <pbonzini@xxxxxxxxxx>: >> On 08/01/2018 10:57, Jan Kiszka wrote: >>>>> Shouldn't this be done on AMD as well, or is the answer "microcode >>>>> update" there? >>>>> >>>>> Jan >>>> Paolo added it, already in linus tree Linux 4.15-rc7 >>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=0cb5b30698fdc8f6b4646012e3acb4ddce430788 >>> Thanks, I didn't noticed this. >> >> Yeah, they were posted in a separate thread. (Actually Ashok Raj's >> patch was mine, :) but I went with Jim's VMX part and attribution >> because he actually tested it with Google's reproducer). > > We cooperate with AMD guys to test the Scrub hardware GPRs against > latest AMD EPYC, we observe the performance of Redis drop ~11% in the > guest (Centos 7.4, 3.10.0-693). So there are no updates installed in the guest (-693 is the version that was released with 7.4)? Or is it the latest CentOS kernel using IBRS? I don't think scrubbing should be particularly expensive. However, filling the RSB can cost over 50 clock cycles. Thanks, Paolo > redis(redis_version:3.2.10) > > Redis command line: > > redis-server > /dev/null 2>&1 & > redis-benchmark -t ping,set,get -d 128 -c 100 -r 1000000 -n 10000000 -q -P 20 > redis-benchmark -t ping,set,get -d 1024 -c 100 -r 1000000 -n 10000000 -q -P 20 > redis-benchmark -t ping,set,get -d 2048 -c 100 -r 1000000 -n 10000000 -q -P 20 > > Regards, > Wanpeng Li >
