Hi Timothy,
I mainly consulted this:
http://public.dhe.ibm.com/software/dw/linux390/perf/network_tuning.pdf
I know network tuning is not easy, there're so many factors. Deep
understanding of network protocols and OS layer is needed to do a good job.
But this case was not so complicated, it is
James,
Have you consulted any specific LinuxONE or Linux on Z network performance
tuning documentation yet? There are many other parameters and settings that
can influence performance.
I learned a lot from your discussion. Thanks.
My question was from a network micro benchmark test which simulated OLTP
workloads generating packet less than 200 bytes typically. And our result
was only half of the x86's. It's easy to re-produce using uperf/netperf by
sending huge amount of small
On Fri, 17 Nov 2017, Alan Altmark wrote:
> So what you're saying, Sebastian, is that (a) fine-grained control over
> what devices interrupt what CPUs is not in the Z architecture, (b) it's
> not in there for a good reason, (c) providing such a capability would
> adversely affect overall system
Alan Altmark wrote:
>It might, however, be an interesting idea to make the ioctl()s in the
>device driver a no-op instead of not being present or generating its own
>errnos. That way folks can turn the knob, feel better, but not see any
>change since "it doesn't get any better than this".
I like
On Nov 17, 2017, at 4:45 PM, Alan Altmark wrote:
> It might, however, be an interesting idea to make the ioctl()s in the
> device driver a no-op instead of not being present or generating its own
> errnos. That way folks can turn the knob, feel better, but not see any
> change since "it doesn't
On Friday, 11/17/2017 at 11:53 GMT, Sebastian Ott
wrote:
> It's a performance tuning knob - a mechanism for device drivers (and
> administrators) to specifiy that they expect an interrupt on the CPU
that they
> used for IO submission.
>
> Since all interrupts are
On Fri, 17 Nov 2017, Timothy Sipples wrote:
> Sebastian Ott wrote:
> >Setting irq affinity is currently not supported on s390.
>
> This platform's I/O architecture is unique. What are the real-world issues
> associated with not being able to set this parameter on this platform? And
> what are the
Sebastian Ott wrote:
>Setting irq affinity is currently not supported on s390.
This platform's I/O architecture is unique. What are the real-world issues
associated with not being able to set this parameter on this platform? And
what are the suggested mitigations?
> Recently we're testing Mellanox 10GbE performance with Ubuntu 17.04 s390x
> on z13. During the test, we found that interrupt affinity cannot be set
> like other platforms.
I/O-related hacks for other platforms are unlikely to work in the same way on
this hardware; the underlying I/O subsystem
Thanks much for the clarification. -James
On Thu, Nov 16, 2017 at 5:20 PM, Sebastian Ott
wrote:
> James,
>
> On Thu, 16 Nov 2017, Jingmin Zhai wrote:
> > Recently we're testing Mellanox 10GbE performance with Ubuntu 17.04 s390x
> > on z13. During the test, we found
James,
On Thu, 16 Nov 2017, Jingmin Zhai wrote:
> Recently we're testing Mellanox 10GbE performance with Ubuntu 17.04 s390x
> on z13. During the test, we found that interrupt affinity cannot be set
> like other platforms.
>
> We stopped the 'irqbalance' service first, then
>
> echo >
Recently we're testing Mellanox 10GbE performance with Ubuntu 17.04 s390x
on z13. During the test, we found that interrupt affinity cannot be set
like other platforms.
We stopped the 'irqbalance' service first, then
echo > /proc/irq//smp_affinity
But get
"echo: write error: Input/output
13 matches
Mail list logo