[Linux-PowerEdge] Intel's Hyper-Threading vs. Linux Iptables/Netfilter Firewall ?

Scott Harrington scotth01 at sns-usa.com
Wed Mar 20 09:34:43 CDT 2013


> The real answer to the original question:  try it both ways.

I concur, you have to test your specific workload using as realistic 
simulation as possible. Dell's BIOS recommendation whitepapers are very 
helpful and appreciated but we all have slightly different needs.

I found that for a key latency metric in our application, we experienced 
twice as high latency on Dell R820 4-socket E5-4650 servers versus what we 
experience on Dell R620 2-socket E5-2690 servers. The difference was 
eliminated by disabling HT on the 4-socket E5-4650s (syscfg 
--logicproc=disabled; reboot).

Similar experiements on 2-socket E5-2690s did not result in any observable 
difference in our latency metrics with HT on versus off. I also tried to 
control for QPI & NUMA effects by testing under numactl --cpunodebind=0 
and =0,1 and =0,1,3 but the HT made the biggest difference.

However it sounds like your workload might be more about interrupt 
handling than mine, which was more about memory, cache, and 
context-swiching effects. You will need to enlist the help of multiple 
other machines to generate realistic DDOS-like network load. Good luck and 
please share your observations.



More information about the Linux-PowerEdge mailing list