[Make-wifi-fast] Flent test hardware

[Bob McMahon]

The realtek is definitely not ideal.   The test needs to watch the ptp
stats to make sure the corrections are stable over the life of the test and
throw out bad results per test equipment introducing too much error.

Qualifying a NIC for use in test equipment is a bit of a pain.  My
rationale is to avoid consumer grade products, rather leverage the work of
engineers that qualify equipment for data centers, i.e. the data center
market is driving the vendor.   I find the INTC server class NICs to be the
best for this so far.

Bob

On Sun, Nov 5, 2017 at 5:57 AM, Pete Heist <peteheist at gmail.com> wrote:

>
> On Nov 5, 2017, at 2:42 AM, Bob McMahon <bob.mcmahon at broadcom.com> wrote:
>
> I have some brix with realtek and run ptpd installed with fedora 25.
> The corrections are in the 25 microsecond range, though there are
> anomalies.  These are used for wifi DUTs that go into RF enclosures.
>
> [root at hera ~]# tail -n 1 /var/log/ptpd2.stats
> 2017-11-04 18:33:46.723476, slv, 0cc47afffea87386(unknown)/1,
> 0.000000000, -0.000018381,  0.000000000, -0.000018463, 1528.032750001, S,
> 0.000000000, 0, -0.000018988, 1403, 1576, 17, -0.000018463,  0.000000000
>
> For LAN/WAN traffic, I tend to use the intel quad server adapters in a
> supermicro mb desktop with 8 or more real cores.  (I think the data center
> class machines are worth it.)
>
>
> Thanks for the info. I was wondering how large the PTP error would be with
> software timestamps, and I see it’s not bad for most purposes.
>
> Which Realtek Linux driver does your brix use, and is it stable? The r8169
> driver’s BQL support was reverted at some point and it doesn’t look like
> that has changed.
>
> I trust that the extra cores can help, particularly for tests with high
> flow counts, but my project budget won’t allow it, and used hardware is too
> much to think about at the moment.
>
> Do you (or anyone) know of any problems with running the Flent client and
> server on the same box? In the case of the Proliant Microserver, the
> Broadcom 5720 adapter should have separate PCI data paths for each NIC. I
> guess the bottleneck will still mainly be the CPU. To get some idea of
> what's possible on my current hardware, I tried running rrul_be_nflows
> tests with the Flent client and server on the same box, through its local
> adapter (with MTU set to 1500) with my current Mac Mini (2.26 GHz Core2 Duo
> P7550). I know that doesn’t predict how it will work over Ethernet, but
> it’s a start.
>
> https://docs.google.com/spreadsheets/d/1MVxGsreiGKNXhfkMIheNFrH_
> GVllFfiH9RU5ws5l_aY/edit#gid=1583696271
>
> Although total throughput is pretty good for a low-end CPU, I’m not sure
> I’d trust the results above 64/64 flows. 256/256 flows was an epic fail,
> but I won’t be doing that kind of test.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.bufferbloat.net/pipermail/make-wifi-fast/attachments/20171105/2a3cba5d/attachment-0001.html>