[OpenIndiana-discuss] zsf send/receive performance

[Jim Klimov]

On 2013-12-09 21:38, wim at vandenberge.us wrote:
> That is an interesting thought. I did check the aggregate CPU utilization (Dual
> 6-core 2.0GHz Xeon, E5-2630L) and overall it looks fine with a load average
> below 0.5 during transfers. However,  I have not checked if there is a single
> thread that is blocking the process. I guess the cheap way to verify your
> assumption is to run the process on a faster processor. If the throughput goes
> up with the speed of the CPU that would indicate we're on the right track. I
> think we have a couple of E-2637v2's laying around, let me see if I can get
> those swapped in.


I think an even cheaper way would be to not use other cores for tasks -
to see if the enforced single-threaded performance would be similar.

For example, you can use "psrset" to define a vCPU-set (made of your
other 11 cores) which would be yanked out of the default set and only
leave one CPU for the OS. Also "psradm" can be used to set vCPU states
such as "offline" so that the scheduler does not put processing threads
on it, and (subject to hardware quirks) interrupts shouldn't come too.

These tests can be done on-line without interrupting your system for
hardware changes or reboots, and you can quickly model different
situations regarding multi-processing (i.e. on systems with SPARC CMT
or Intel HyperThreading you can create sets that would show performance
differences of cores vs. threads/"strands" as parallel calculators).

HTH,
//Jim Klimov