[openib-general] Re: BUG ipoib oops in iptables/netfilter

[Grant Grundler]

On Wed, Mar 30, 2005 at 11:08:31AM +0200, Michael S. Tsirkin wrote:
> Quoting r. Grant Grundler <iod00d at hp.com>:
> > FWIW, default IPoIB perf is pathetic: ~1.5-1.6Gb/s
> > with the above netperf command line.
> 
> I think these systems have IOMMU, do they not?

They do. But HP ZX1 chipset allows 64-bit devices to bypass
the IOMMU and directly address memory.

tg3 is a 64-bit device and bypasses the IOMMU.
ie tg3 is using physical addresses.

> If so, could the fact that IPoIB calls dma_map_single for each
> packet be the reason it is slow? Maybe dma_map_single is slow?

pfmon output I've collected in the past doesn't support that theory.
The ia64 "machvec" indirect function call certainly isn't helping though.

ZX1 is about 3yr old chipset.
I expect current x86/amd64 chipsets are not as old.
I really need to collect PCI bus traces to verify if it's
a chipset problem though.
The logistics of doing that are non-trivial and I'll
need help from a local HW group. *sigh*

thanks,
grant