This is a common recommendation really with hpc applications,
particularly when doing higher-bandwidth operations, such as networking
at 10-100gb interfaces. Hyperthreading arbitrator in the kernel is like
a buffer, the L1-2 cache (I think), that fills as the cpu backs up.
When full, the arbitrator can't pass to the cpu, sits in buffer, and
eventually gets there (hopefully). This is BAD when you are doing very
latency sensitive crunching.
Likewise, things like irq balancing are generally disabled for the same
reason to keep hardware like ethernet and drive hba's stable and
low-latency. You try to design to the cpu workload, memory, pci
bandwidth, ethernet hardware, etc, to not *need* buffers, at least when
a dev understands such things, which is generally few and far between.
-mb
On 09/07/2016 01:27 PM, Kevin Fries wrote:
>
> I once worked for a company doing ground water modeling for mining
> operations. The program did a large series of fourier transformation
> to model the water levels over time... No Hyperthreading!!!
>
> Most web servers, mail servers, database servers (depending on your
> number of indexes), are perfectly fine with hyperthreading turned on.
>
> Kevin
>
---------------------------------------------------
PLUG-discuss mailing list -
PLUG-discuss@lists.phxlinux.org
To subscribe, unsubscribe, or to change your mail settings:
http://lists.phxlinux.org/mailman/listinfo/plug-discuss
(text/plain)